use case library
If the process
is known, an agent can probably run it
Exposing internal systems to agents without exposing your data
Connecting an agent to an internal system is easy. Connecting it so it sees only what the job needs, and cannot act beyond it, takes design. This page sets out that design: narrow views, separate accounts, read before write and a person on every consequential action.
Making a WordPress site readable by agents
Key facts on WordPress sites often sit in images, sliders and scripts that AI agents read poorly. An agent drafts the fixes in WordPress as drafts, page by page, and the site owner reviews and publishes every one.
Is your website agent-ready
AI assistants read websites on someone’s behalf, and what they see can differ from what a person sees. An agent audits your site the way an agent reads it and writes up what it finds. Every change is the site owner’s decision.
Who approves a model going to production
When a model goes live, somebody is accountable for it, whether or not anyone wrote that down. This page sets out who that approver should be, what they need in front of them, and how the decision gets recorded.
Eval harness runs on every model change
Teams change prompts and models often and check the results by trying a few examples. An agent runs the whole evaluation suite on every change, compares it with the last approved run and writes the report. Whether the change ships is a person’s call.
The agent-run managed services back office
A managed services business runs the same back office for every client: tickets, monitoring, onboarding, licenses, billing and the quarterly report. This is the pattern for running it with governed agents, each under a named person, with every client commitment kept human.
The NOC coordinator as an AI Employee
In a small operations team, the engineer fixing the outage is also the one answering the phone, updating the ticket and writing the handoff. An AI Employee takes the coordination as one supervised job, and every operational decision stays with engineers.
Asset and license inventory reconciliation
The asset spreadsheet says one thing, device management says another, and the software bill says a third. An agent compares them on a schedule, lists every mismatch with its evidence, and hands each one to its owner. Reclaiming, buying and retiring stay with people.
Ticket triage and escalation across queues
IT tickets land in the general queue, wait for someone to notice they belong to infrastructure, and the urgent one sits next to a password reset. An agent sorts every ticket by your rules, escalates the urgent ones and shows its reasoning. Engineers fix.
Change request intake and change board packet preparation
Change requests arrive missing the rollback plan, and the change board spends its meeting asking for it. An agent checks each request against your required fields, gathers the context and builds the packet. The board approves or refuses every change.
Vulnerability and dependency monitoring with triage
Vulnerability alerts keep arriving and sit unread, including the few that matter. An agent gathers them, checks which are known to be exploited and which systems they touch, and hands each engineer a triage note. Engineers decide and fix.
Incident timeline and postmortem assembly
After the outage, someone spends an afternoon scrolling chat to work out what happened when. An agent assembles the timeline from the records as the incident runs and drafts the postmortem. Causes, findings and follow-ups are the incident lead’s to decide.
Runbook execution with a human gate on destructive steps
Runbooks are written down, then run by hand in the middle of the night, because nobody trusts automation with the step that deletes something. An agent runs the safe steps and stops at every destructive one until a named person approves it.
Access provisioning and approval routing
Access requests arrive by chat, get granted by whoever is around, and nobody remembers who approved what. An agent records each request, routes it to the system owner, and passes approved ones to IT. Approving and granting stay with people.
Why AI drafts and attorneys decide
AI can draft, compare, check and organize legal work. What it cannot carry is the professional duty attached to advice and to a signature. That is why the line sits where it does: AI drafts, attorneys decide.
The contracts administrator as an AI Employee
Contract administration is intake, routing, templates, calendars and filing, and at a growing company it falls on a lawyer’s evenings. An AI Employee takes that administration as one supervised job, and every legal judgment stays with lawyers.
Research briefs where every citation gets verified before it ships
A research brief is only as good as its citations, and a citation nobody checked is a risk with the firm’s name on it. An agent drafts the brief and verifies every authority against the source, and the attorney reviews, decides and signs.
Matter status reporting clients can read
Clients want to know where their matter stands, and the answer lives in a docket and a task list written for lawyers. An agent drafts a plain-language report from that record, and the responsible attorney edits, approves and sends it.
NDA turnaround on standard terms, escalation on everything else
A routine NDA request waits in legal’s queue behind larger work. An agent prepares your standard NDA, routes it to the person who signs, and sends anything non-standard to a lawyer with the changes marked.
Obligation extraction and deadline calendaring
Signed agreements promise reports, notices, audits and deliverables, and most of those promises live only in the PDF. An agent lists each obligation with its clause quoted, a lawyer confirms the list, and every item gets an owner and a date.
First-pass contract review against your playbook
Most of a first read is checking the same clauses against the same standard positions. An agent does that comparison against your written playbook and lists every departure with both texts side by side. Interpretation, negotiation and the decision stay with the lawyer.
Contract intake, classification, and routing
Contracts arrive from every direction, and the first hour of legal’s day goes to working out what each one is and who should look at it. An agent logs, classifies and routes each one to the right reviewer. Reading, judging and signing stay with people.
Headcount and turnover reporting that builds itself
Every quarter someone rebuilds the headcount spreadsheet, and the turnover number changes depending on who built it. An agent produces the report from your HR records with definitions you wrote down, and a named person reviews it before anyone else sees it.
The HR coordinator as an AI Employee
The HR coordinator’s week is scheduling, chasing, answering and reporting, and it lands on whoever has a spare hour. An AI Employee takes that role as one governed job, reporting to a named person, while every decision about people stays with people.
Open enrollment support without the HR inbox pileup
For a few weeks a year the same benefits questions flood one inbox, and the deadline sneaks up on the people who never opened the email. An agent answers from your plan documents, tracks completion and reminds people. Plan choices and eligibility stay with humans.
Offboarding as one checklist instead of four
When someone leaves, four teams each keep their own list, and the account nobody remembered stays open for months. An agent turns the departure into one checklist with owners and dates. Revoking access, final pay and the conversation stay with people.
Onboarding coordination across IT, payroll, and facilities
The offer is signed, and now four teams each need to do something before day one, and nobody is watching all four. An agent opens every setup request with an owner and a date, chases the stragglers, and reports what is ready. Access and document checks stay with people.
Employee policy questions answered from your own handbook
The handbook has the answer, but nobody reads it, so the same questions land on one person every week. An agent answers from the handbook with the section quoted, says when it is silent, and hands anything sensitive straight to a person.
Performance review assembly from the work that actually happened
Reviews get written from the last few weeks and whatever the manager remembers. An agent assembles the cycle’s actual record, every line linked to its source, and stops there. The rating, the words and the conversation stay with the manager.
Interview scheduling and candidate communication
Interviews get scheduled by email tennis, and candidates hear nothing for a week. An agent books slots against real calendars and sends the updates you approved, while who advances and who is hired stays with the people doing the hiring.
Engagement letter and e-file packet prep
Every season the same letters get retyped and the same signature packets get assembled by hand. An agent drafts both from your templates and the client record, and the named preparer or attorney reviews and sends them. The client signs, and the preparer files.
Prospect list building from written criteria
Prospect lists get built from a vague profile and whatever a rep found on a Tuesday. An agent builds the list from criteria you wrote down, shows the evidence for every match, and contacts no one until a person approves the list.
Travel booking and itinerary coordination
Booking a trip means a pile of browser tabs, a policy nobody rereads and an itinerary pasted together the night before. An agent assembles options inside your policy and builds the itinerary, and the traveler or approver confirms every booking and every payment.
Website and CMS content updates
Small website changes wait weeks behind whoever holds the CMS login. An agent takes each request to a saved draft with the before and after side by side, and a named reviewer decides what goes live.
The agent-run finance back office
Nine finance workflows, each proven on its own page, become one finance function under one set of oversight roles, one exception owner and one audit trail, so work crosses between them on the record. This page describes that pattern and names who holds each gate.
Governed development: how our agents ship work orders through review gates
Agents write code and close work before anyone reads it. Governed development is the discipline that answers for it: scope fixed in a work order before code, separate lanes to build and to check, evidence carried with every hand-off, and a named person who approves what ships.
Building on undocumented and esoteric third-party systems
Your product depends on a system nobody documented, and every guess about it is a hidden risk. This is how an agent maps what it actually does, keeps the observed apart from the assumed, and contains it behind one adapter in your code, while the third-party system itself is never changed.
What an AI-built codebase is missing (the assessment checklist)
The product works, but nobody can say which parts are safe to touch. This read-only assessment lists what an AI-built codebase lacks, with the evidence for each finding, and changes nothing: your product owner decides what happens next.
Hardening a vibe-coded MVP without a rewrite
The product has users, and every change breaks something nobody expected. This is the repair-in-place path: pin what the paths users depend on do today, then propose one small change at a time, each read and merged or sent back by a named reviewer on your side.
The compliance audit that assembles itself
The package gets built in the week before, out of folders nobody has opened since the last one. This is how an agent holds the scope your own record defines and keeps the package standing against it, with the source of every item shown and the review, the signature and the opinion left to people.
Hiring without an HR department: you decide, the bias audit writes itself
The criteria live in your head, and applications get read in a different order by whoever has the hour. This is how an agent applies the criteria you wrote down to each one, keeps the record of what was applied to which as the screening runs, and leaves every decision about a person to you.
Incident documentation and reporting
The incident gets written up weeks later, from three people’s memories of it. This is how an agent opens the record the moment somebody raises one, assembles from what your authorized records already hold, and leaves reportability, notification and closure to the accountable owner.
Certification renewal management
The certification lapsed because the renewal needed something nobody had collected yet. This is how an agent holds the register you keep, watches each date against the lead time you wrote down, and assembles what your own record says each renewal requires, while the application stays with a person.
Policy adherence monitoring with exception flags
The policy was approved once and nobody has checked it since, so adherence is assumed until somebody asks. This is how an agent runs the comparison continuously, binds every exception to the rule it touched, and holds the question of whether it is a violation for the person accountable.
Regulatory deadline tracking and evidence assembly
The register lives in one person’s calendar and the evidence lives in folders nobody opens until the week before. This is how an agent holds both together through the period, while every filing and attestation stays with the person who signs it.
Monthly client billing on autopilot, with sign-off on every run
The billing run gets re-specified from memory every cycle. This is one written definition per client, one run assembled against it, a line-by-line comparison with the prior cycle, and a signature from the named billing owner on every run before anything leaves.
Audit documentation that stays ready year-round
The support file gets rebuilt every time a request arrives, from records that existed all along. This is collection running alongside the work, filed against the requirement it answers, with the pack and the gaps routed to a named owner before anything leaves.
Budget variance monitoring and alerts
A variance arrives as an alert with no story attached, long after the spending happened. This is the comparison against the lines you set, on the cadence you set, with the apparent cause attached and the breach held open until the named owner of that line closes it.
Expense report processing and policy checks
An approver reads a whole report to find the one line that breaks policy. This is the check run per line against the policy you wrote down, with the failing rule named beside the failing line, and approval, reimbursement and payment left with a person.
Cash position reporting that builds itself
The cash view gets rebuilt by hand before every decision. This is the position assembled on the cadence you set, from balances you already hold, with stale sources named as stale and the named finance owner reading it before it is circulated or acted on.
Cross-system reconciliation on a daily cycle
Two systems hold the same fact and disagree by the end of the month. This is the comparison run on the cadence you set, on the keys your rules name, with both sides of every difference shown and the decision routed to the named finance owner.
AR follow-up that never forgets
Follow-up fails because nobody had the invoice on a list that week. This is the sequence written down once, run against invoices you already issued, with every message past the tier you authorized held for the named finance owner.
FAQ and help content that updates itself
The help centre was accurate when it was written, and nobody finds out otherwise until a customer says so. This is how an agent starts from the questions people actually asked, compares the published answer against your own records, and publishes nothing itself.
Voice-of-customer synthesis from tickets and calls
Everybody half knows what customers keep saying, and nobody has written it down. This is how an agent groups the records you authorize into your own themes, keeps every theme traceable, and asserts no outcome.
Renewal and churn-risk monitoring
The renewal that surprised you was visible for months in records you already had, and looking was nobody’s scheduled job. This is how an agent watches your own written signals against your own written thresholds, and decides nothing.
Customer onboarding coordination
The onboarding plan lives in a document and stops being followed once the work starts. This is how an agent holds the plan with an owner on every step, drafts the chase for a person to send, and commits nothing on a customer’s behalf.
Support ticket triage and routing
A ticket sits because nobody has decided what kind of thing it is, and the person who could answer it does not know it is waiting. This is how an agent triages, routes and drafts, while every send and every ticket write stays with a person.
Turning webinars and calls into a content library and meeting recaps
The best answer anybody gave sits deep inside a recording nobody opens, and the action items from last week’s call sit in nobody’s inbox. An agent drafts assets and recaps from recordings you own, shows where every line came from, and leaves sending to a person.
Social and channel data unified into one report
Three exports, three numbers that claim to be the same thing, and a reconciliation that lives in one person’s spreadsheet. This is how an agent records the mapping onto your own definitions, names what must not be summed across channels, and states a late period as late.
How to monitor how your brand shows up in AI search answers
How can you monitor how your brand shows up in AI search answers? Keep a fixed list of buyer questions, run them on a consistent set of answer engines, and save every answer with its date, citations and exact brand mention or recommendation. Compare observations by the same promp
Competitive monitoring without the tab farm
The watch is real work nobody has written down, and it lives in one person’s browser. This is how an agent holds the list, the cadence and the categories, reports what changed against the last check with its source, and says when a source could not be reached.
Lead scoring beyond gut feel
The criteria for a good lead exist, and they live in somebody’s head. This is how an agent scores from criteria you wrote down, shows the inputs behind every number, and names the record your criteria cannot score instead of scoring it low.
Content production pipelines that ship on schedule
A piece slips because its stage, its owner and the claim that still needs a source are written down nowhere anyone can look. An agent holds all three, drafts with each claim’s source shown, queues a social post only after a named person approves it, and leaves the newsletter send to a person.
The governed research brief
The answer came back confident and you cannot check it. This is the pattern for a research brief whose statements trace to sources and whose last step names what the evidence does not support.
Knowledge base maintenance that doesn’t rot
An article goes stale because it has no owner and no date by which anyone was going to look at it again. This is how an agent holds both, watches the sources, and publishes nothing itself.
Onboarding paperwork and checklist coordination
The checklist exists and so does the policy behind it, but both live in somebody’s memory of the last joiner. This is how an agent runs them as written and determines nothing about employment.
Status meetings replaced by status agents
The standing call exists because nobody trusts the record enough to skip it. This is how an agent assembles the status and each accountable owner validates their own section, with which meetings stay left to you.
Meeting scheduling and calendar coordination
Every scheduling thread re-derives the same rules about your time by hand. This is how an agent holds those rules in one place, proposes against them, and writes to no calendar at all.
Internal request routing and triage
A request from a person arrives in whichever channel was nearest and missing the one field anyone needed. This is how an agent classifies, completes and routes it, and fulfils nothing.
Contract deadline and renewal tracking
The notice window sits inside a document that was filed correctly, and nobody reads it again until it has closed. This is how an agent holds the dates and the owners while the decision stays yours.
Sales follow-up that keeps the thread
Threads go quiet while everyone assumes somebody else has it. This is how an agent keeps a ledger of what is open, flags the thread that has gone quiet against the interval you set, and hands a person the context and a draft.
Deal desk approval routing
The exception is the part of a deal that actually needs a decision, and it usually gets made in a chat thread. This is how an agent assembles the request against your policy and keeps it visible, while the approval stays a person’s every time.
Forecast assembly your CFO will trust
A forecast submission is a set of figures plus the working behind them, and the working is what goes missing. This is how an agent assembles the pack from your records and your method while the judgement and the committed number stay a person’s.
Follow-up sequencing that adapts to replies
Your follow-up rules exist, in somebody’s head, so what happens after a reply depends on who read it. This is how an agent classifies the reply against written categories and prepares the next step, with the send left to a person every time.
Security questionnaire automation
The questionnaire arrives mid-deal and the answers get rebuilt from the last submission. This is how an agent drafts them from your own maintained library and marks what it cannot support, with a person verifying every answer before anything goes back.
Meeting prep briefs before every sales call
The call is on the calendar and the context is spread across a record, a thread and a meeting nobody wrote up. This is how an agent assembles one brief per call from what you already hold, while the conversation stays a person’s.
Pipeline reporting without the Friday spreadsheet
The weekly pipeline view gets rebuilt by hand from records that disagree with each other. This is how an agent assembles it from those records, names the disagreements instead of smoothing them, and stops before the commentary and the release.
Proposal and quote assembly from your templates
The proposal that goes out is usually the last one with the names changed. This is how an agent assembles it from your current template and your own record, shows where every value came from, and leaves the price, the terms and the send to a person.
Recurring reports that spawn themselves
Every recurring report is rebuilt as if it were new, by whoever remembers it is due. This is how an agent instantiates each cycle from a definition written once, with sign-off by the owner before anything is distributed or used.
Vendor negotiation prep and coordination
The negotiation is lost before the meeting, in the scramble to find the contract and in the stakeholder nobody asked. This is how an agent prepares and coordinates the internal side while the negotiation and every vendor contact stay a person’s.
The agent-run operations back office
Mercury Digital builds custom AI agents that run back-office work for mid-size businesses: close, AP, approvals and reporting, with named human oversight.
Scheduling and dispatch coordination
The schedule is rebuilt each morning from the same inputs by the same person, and every change after work starts is a phone call. This is how an agent proposes and records while the dispatch decision and every field contact stay the dispatcher’s.
How AI produces continuous project status reports from existing tools
How can AI produce continuous project status reports from existing tools? Give an agent authorized read access to the plan of record, task or ticket records, milestones and approved project notes, then have it assemble a status draft on a set cadence. Every statement should trace
Supplier scorecards unified across systems
Every system holds one piece of how a supplier is doing, and the scorecard exists only in the meeting where someone argues from memory. This is how an agent assembles it on a cadence while every supplier action and contact stays a person’s.
Inventory monitoring and reorder triggers
The threshold is written down and the level is recorded, and nobody compares them until something runs out. An agent runs the comparison on a cadence, for warehouse stock or the supply closet, and stops at a reorder a person approves.
Purchase approval workflows that don’t stall
An approval stalls in the gap between the request and the decision, where nobody owns the chase. This is how an agent assembles, routes and tracks each request while the approval, the order and the payment stay a person’s.
Exception detection and routing across your ops stack
The exception already exists in one of your systems; nobody is watching the seam between them, so it is found by the person it hurts. This is how an agent detects, classifies and routes while the correction in any system stays a person’s.
When the system notices your busywork before you do
The work that eats your time is rarely on any list, because nobody decided to do it. A governed work record can show the pattern. What the system does with it is propose, and a named person decides.
Proactive drafting from email & tickets: the work order you didn’t create
Work arrives as messages. The useful pattern is an agent that reads a named intake source and drafts a work order that is inert until a person promotes it: no reply, no dispatch, no owner.
Plans vs routines: projects & recurring processes, governed differently
You govern a migration and a monthly billing run with the same approval habit. One will be done and the other never will. Plans and routines are two shapes of governed work, and the shape decides the oversight.
When NOT to automate: workflows that should stay human
A process resists a handoff for one of two reasons: it is not ready, which preparation fixes, or it should not move, which is a property of the work. Sorting the two is the job.
What an agent-run department actually looks like
The phrase covers everything from one chatbot to an org chart with no people. The shape is several agents on one function, a management layer holding the record, and named people holding the gates and the exceptions.
Shadow mode: how agents earn autonomy
Shadow mode runs an agent beside your team, touching no real record, until a named person decides it has earned a gate. Checklist and exit criteria.
Approval gates: how to keep humans in the loop without slowing down
An approval gate puts a named person’s sign-off on the step that matters. How Mercury Digital places gates, what they cost, and how they map to NIST AI RMF.
Turning a job posting into an AI Employee spec
A posting for a role that is mostly process already describes a system. The method sorts its lines into known-process work and judgement work, and turns the process lines into specified, gated workflows.
The compliance coordinator as an AI Employee
Compliance coordination is a calendar, an evidence trail and a policy watch that only work if they are kept together. This is how an agent keeps them together while the accountable person keeps every signature.
The executive assistant layer as an AI Employee
The assistant’s job is context rather than typing. This is how an agent holds the schedule, the prep and the follow-through for one executive while every commitment made in their name stays with them.
What an AI employee for marketing operations actually does
An **AI employee for marketing operations** coordinates recurring work across your campaign calendar, metric definitions, channel reports, content pipeline and authorized website changes. It assembles records and drafts, checks them against your rules, and routes spend, launches,
The sales ops coordinator as an AI Employee
Records, reports and proposals share one dependency, the record. This is how an agent holds the coordinator’s seat between them while the revenue owner keeps every decision that reaches a customer.
The AR clerk as an AI Employee
Follow-up, payment tracking and the receivables report are one role on paper and three interruptions in practice. This is how an agent holds the seam between them while a person decides what leaves the business.
How to decide which admin workflows to automate with AI first
How can you decide which back-office workflows to automate with AI first? Start with recurring work that has named inputs, written decision rules, a checkable output and an owner who can review exceptions. Good first AI automation projects for a mid-size company’s operations team
How companies use AI for document intake and PDF extraction
How do companies use AI for document intake and data extraction from PDFs? An agent captures files from approved intake channels, classifies each document, extracts only the fields required for that type, validates the extraction against the source and routes low-confidence cases
Vendor onboarding without the email ping-pong
Onboarding stalls in the gaps between requests, whether the party owing paperwork is a vendor or your own client. An agent owns the chase and the status, and a named person owns every approval and every review of what arrives.
Campaign reporting that assembles itself
The monthly report is the same joins rebuilt every cycle. Move the assembly into a continuous job and the meeting changes from reconciling numbers to deciding what to do about them.
Lead qualification and routing in minutes
Routing fails at the handoff more often than at the scoring. This is how an agent qualifies against written criteria, delivers context with the lead, and gates what it cannot decide.
Keeping CRM data clean without nagging reps
Hygiene fails because it is a tax charged to the person least willing to pay it. This is how an agent maintains the record from activity your team has already generated.
Invoice intake and exception triage
How Mercury Digital’s agents read, match and code invoices, sort exceptions by type, and route every payment decision to a named approver.
Automating month-end close
How Mercury Digital’s agents reconcile through the period, draft entries with support attached, and route every posting to your controller for sign-off.
