The work that eats your team's week.
Six patterns cover most of what we're asked to build. Yours is probably a variation of one of them — with the numbers we would hold ourselves to, and the decision we deliberately leave with a person.
Your analysts are a very expensive copy-paste layer.
An analyst gets 200 target companies on Monday. By Wednesday they have done sixty. On Thursday the list changes.
An agent works the list: pulls from sources you already trust, cross-checks every figure against a second source, and writes into your schema. Where it cannot verify something it says so instead of guessing — the difference between a research tool and a liability.
- 31 → 2analyst hrs per week
- 2 days → 20 minrequest to delivered list
- 2 sourcesevery figure cross-checked
- flaggednever guessed
Agreed in writing before the build starts.
The analyst still owns the conclusion. Anything the agent could not verify is surfaced at the top, not buried.
The deliverable you produce for the fortieth time.
Every proposal starts as last quarter’s proposal. Someone strips the client name, rewrites four sections, and misses a figure in the appendix.
A generator that reads your own template library and data, assembles the first draft in the right structure, and pulls live figures from the system of record instead of from the last document.
- hours → minutesbrief to first draft
- fewer roundsbefore sign-off
- live figuresnot last quarter’s
- your voicetrained on your work
Agreed in writing before the build starts.
It produces a draft, never a send. The recommendation, the pricing and the commitment stay with the person whose name is on it.
Four hundred applications and one afternoon.
Four hundred applications. The first sixty get real attention and the rest get thirty seconds each.
Scoring against your rubric, not a generic one, with a written justification per item that a person can disagree with. Everything arrives ranked, borderline cases at the top rather than the bottom.
- all of itshare that gets a real read
- rankedwith written rationale
- hours → minutesper hundred screened
- every onerejections seen by a human
Agreed in writing before the build starts.
A human approves every rejection. An automated rejection you cannot justify is a legal exposure, and every candidate experience is your brand.
Eleven emails to book one meeting.
Three partners, two time zones, one candidate, and a coordinator with six calendars open. Offer three slots, two get taken by client work, start again.
An agent that holds the constraints — seniority, panel composition, time zones, buffer rules — finds the overlap, books it, and sends the brief. When someone drops out it re-solves rather than escalating.
- 11 emails → 0to book one meeting
- same dayready-to-booked
- autoreschedules handled
- no coordinatoropening a calendar
Agreed in writing before the build starts.
Who sits on the panel stays yours. The agent never moves a client meeting to solve an internal one.
The month-end work nobody wants to own.
It lands on whoever is least able to say no. Invoices keyed twice, timesheets chased by Slack, a compliance checklist somebody remembers to open.
Extraction and matching against your PO and project structure, coded to the right cost centre, with a queue that surfaces only the exceptions. Chasing happens on a schedule, so nobody has to be the person who nags.
- days → hoursto close the month
- exceptions onlyreach a human
- on scheduletimesheet chasing
- per invoicecost tracked
Agreed in writing before the build starts.
Anything that moves money stays a decision. Payment, write-off and vendor approval keep their evidence attached so the call takes seconds.
The same report, every Monday, forever.
Someone spends Friday afternoon pulling the same numbers out of the same four systems into the same deck. If they are on leave, the client notices.
A scheduled assembly from your systems of record into the client’s template, with the movements since last period called out rather than left for the reader to spot.
- before stand-uplands every time
- on leavestill goes out
- called outwhat actually moved
- one personcan carry them all
Agreed in writing before the build starts.
The commentary and the send stay human. A partner decides what the numbers mean before a client sees them.
AI hiring automation, worked all the way through.
Hiring is where several of the patterns above land at once, so it's the example we can be most concrete about. This is one use case, not our whole business — the same four workflows have equivalents in every function that runs on people's hours.
Résumé screening with competency scoring
Scored against your own consulting rubric — structured reasoning, domain depth, communication — with a written justification per candidate. The output is a ranked shortlist a hiring partner can argue with, which is the only kind worth having.
Interview scheduling across partner calendars
The highest-frustration, lowest-risk, most visible win in the whole function. Panel composition, seniority rules, time zones and buffers held as constraints, re-solved automatically when someone drops.
JD and role-kit generation
A hiring manager's five-line brief becomes a job description, screening questions, a scoring rubric and outreach copy — in your firm's voice, ready to edit rather than ready to send.
Hiring-funnel reporting
Assembled from the ATS on a schedule: stage conversion, time-to-shortlist, offer-to-join drop-off, agency spend. The numbers that tell you whether any of the above is working.
The system scores, ranks and explains. It never sends a rejection, and it never makes an offer. This is both the ethical line and the practical one: an automated rejection you cannot justify in writing is a legal exposure, and in a firm that hires from a small number of campuses, every candidate experience is your brand. Candidate PII stays inside your tenancy throughout — the architecture is here.
However we build it, you choose how you hold it.
The same system, four ways to own it. You can start in one column and move to another later — switching is not a renegotiation.
| Aspect | Hosted API | Embedded in your environment | Buy the source code | Buy the code, keep the team |
|---|---|---|---|---|
| Who hosts it | We do | You do | You do | You do |
| Where your data lives | Our infrastructure, region-pinned | Inside your cloud or VPC, never leaves | Wherever you deploy it | Wherever you deploy it |
| Who can modify it | Us | Us, with your access | You, or anyone you hire | You and us, together |
| What you own | The outputs and your data | The outputs, your data, the deployment | Full IP in the source, documented | Full IP, plus the people who wrote it |
| Ongoing cost shape | Usage-based, capped monthly | Your infra cost plus your model usage | Your infra and model usage only | The above, plus a separate service line |
| How you exit | Stop calling the API | Buy the code, or switch it off | Nothing to exit — it's yours | Cancel the service, keep the code |
Scroll the table sideways to compare all four modes.
Which of these is your Tuesday?
That's the audit. Forty-five minutes with your team, then a written shortlist of what's worth automating in your firm — ranked by hours saved against effort to build. You keep it either way.