What Else Can It Do?

Every owner we talk to asks the same question in the first ten minutes, and it is a good one. Apart from answering my questions, what else can it actually do?

The honest answer has changed twice in eighteen months, and the change has not reached most desks. In 2024 the answer was drafts. In 2025 it was drafts and lookups and a first pass at the mail. In 2026 the answer is that it can hold a desk: the front desk, the invoice chasing, the file, the customer replies, the first version of the software your firm runs on. Not help with the job. Do the job, under rules, with a person checking.

That is a bigger claim than “it writes well”, so this piece is built to be checked. Every desk below either carries a named case with a figure and a source, or says plainly that we could not source one yet. Nothing is filled in by guessing.

Last year it answered. This year it holds a desk.

The thing on your phone that answers questions and the thing that runs a front desk are the same model. What differs is whether somebody set it up to do a job: gave it the calendar, told it in writing what it may say and what it must never say, and put a person where the decisions are. Left alone it waits for a question. Set up, it works.

The gap is measured. In May 2026 the US Chamber of Commerce Foundation asked 1,070 people who work at small businesses what they use AI for. Sixty-four per cent said personal productivity — drafting, summarising, brainstorming. Six per cent said workflows that run with minimal human involvement.⁷ Half of small-business workers use it. One in sixteen has it doing a job.

We know the other answer because it is how this company runs. The site you are reading, the campaign platform launching in October, and the four systems on our record page were built by AI working under written rules, with one person deciding and checking.¹ The engineer’s desk and the product manager’s desk in the figure below are not a forecast. They are the two desks we handed over first, and the record of every time that went wrong is public.

So here are the ten. Drag the control, or let it run.

The ten desks

The front desk. The phone gets answered, the appointment gets booked and confirmed, the message gets taken and passed on. This is the desk most small practices hand over first, because it is the one that rings at 6pm. Wolfe Dental, two locations near Portland with one receptionist each, put an AI voice agent on overflow and after-hours calls in August 2025. Missed calls went from about 500 a month to around 50, and the practice reports $134,000 of production booked between September and April.⁸ That is the vendor’s own case study, unaudited, and production booked is not cash collected. The calls it cannot handle — complex billing, clinical questions — it hands to the desk, and the office manager is quoted by name.

Chasing invoices. When to send the first reminder, how the second differs from the first, which customers are never chased because you would rather lose the invoice than the relationship. The previous entry walked through this desk in detail: the software gets the structured facts — amount, date, terms — and never the customer’s own remittance notes, because those are somebody else’s words.² It drafts; a person signs the ones over a threshold.

In the UK it has gone further. A one-person HR consultancy chased a hospitality client’s unpaid fees through Garfield AI, the first AI law firm regulated by the solicitors’ regulator, and won a £7,000 award at Wandsworth County Court in May 2026 for about £400 in fees, with a human barrister arguing the three-hour trial.⁹ The regulator’s condition is the one this piece keeps returning to: it will not take a step the client has not approved. Its ladder is priced per rung — £2 for a polite chaser, £7.50 for a letter before action — and its own example of how the second letter differs from the first is that doctors always send the polite one.¹⁰ Who is never chased stays the owner’s list. No source we found lets the software decide that, and none should.

Keeping the file straight. Every call and email that changes a client’s situation gets written into the record the same day, in the same place, so the next person who opens the file sees what the last person knew. The failure mode of every small firm is not a missing record. It is a record that was true in March. NisonCo, a ten-person PR agency in New Jersey, has an agent read the transcript of each client call, pull out the commitments, and write them into the firm’s CRM the same day; a second agent drafts the follow-up email and leaves it in the founder’s drafts folder, unsent, for him to read.¹¹ Zapier’s own customer story, and the founder’s own estimate of what the call agent is worth is “probably another $500 a month at least”. The machine writes; the owner sends.

First drafts. Quotes, proposals, notices, the letter nobody wants to write. It writes the first version from the file and the template; the owner edits and sends. Industrialized Construction Group, a five-person construction consultancy, has its past proposals pulled into a template so a new one starts as a draft instead of a blank page; the partners report proposal response time down 80 per cent, in Microsoft’s own customer story, with no baseline stated.¹² Earley Law Group, a nineteen-person injury firm in Massachusetts, has its demand letters — the long, fact-heavy letters that open settlement talks — drafted from the medical file in ten to fifteen minutes instead of hours, in the vendor’s case study.¹³ Neither page says who reads the draft before it goes out. This piece does: the owner, every time, until the count says otherwise.

Customer service. The routine questions — hours, status, how do I, where is my — answered at once, in your voice, from your own material. The part that decides whether this desk is safe is what happens when it does not know. Sierra, which runs this desk for large companies, publishes the number: asked the same thing eight times running, the 2024 models got it right about one time in four.³ So the desk is built to stop and hand over, not to guess. The handing over is the product. The same shape at a smaller desk: across 65 UK dental practices, a vendor reports 96 per cent of calls resolved by the AI receptionist and 4 per cent handed to staff, with the practices’ own workflows deciding when a person is needed.¹⁴ Hold the numbers loosely — it is the vendor’s release, and no practice is named — and keep the shape. The 4 per cent is the product.

Order tracking. Every job followed to its next step, and a question raised when one stalls: the part that has not arrived, the approval nobody gave, the customer who went quiet. Leonard Splaine Co., a family heating and electrical contractor in Virginia with 26 people on its team page, has software watch every estimate that was given but not sold and keep making contact until the job moves one way or the other. The contractor reports follow-up sales on unsold estimates up 25 per cent since go-live.¹⁵ The service reps stay. The general manager’s words: “I don’t really want to come in here and say, ‘I don’t need CSRs anymore.’

Scheduling. Who goes where, when, with what, and what happens when Tuesday falls apart. Quality Service Company, a South Carolina heating, plumbing and electrical contractor — a bigger firm than this piece is written for, and it should be said — reports that nearly 30 per cent of its bookings now run from the call through to a scheduled, dispatched job with no person in the chain.¹⁶ One firm’s own number, in the vendor’s release, with no baseline and no period. An existence proof, not a benchmark, and the one desk on this list where we could not find a firm of eight doing it yet.

The books. Receipts categorised as they arrive, the bank reconciled, the month closed without a weekend lost to it. A Stanford and MIT working paper, reported by the Journal of Accountancy, compared accountants who let software categorise and draft the books with those who did not, across the ledgers of 79 small and mid-sized businesses: the month closed seven and a half days sooner, and the books got finer, not coarser — 12 per cent more accounts in use.¹⁷ Observational, not causal, and the ledgers came from the software vendor’s own clients. Your accountant still signs the year, and the paper carries its own warning: in a framed experiment, accountants sometimes over-relied on wrong suggestions. Which is the checker’s desk, again.

The engineer. This is the desk that surprises owners most, because it is the one they assumed was furthest away. The largest AI application companies today sell to people who write software: Cursor went from a $1bn to a $4bn revenue run rate in seven months doing exactly that.⁴ And it is the desk we filled first. Every system on our record page was written by AI under a written specification, reviewed by a second AI from a rival lab whose only job was to break it, and released by a person.¹ A firm of eight does not need an engineer on the payroll to have software of its own. It needs the rules the engineer works under.

And it is not only us. Jean-Christophe Arnaud runs Malorian, a five-person event agency in France, and has no coding background. He built the application that now runs the agency — quotes to invoices, margin per job, supplier quotes read by the software — over a weekend, released it to his team, and cut the features they found useless: €2,000 a year of subscriptions replaced for a stated one-time cost of €1,000, in the platform’s own case study.¹⁸ A five-person roofing firm in Ohio did the same with five subscriptions that cost up to $1,800 a month, in the same platform’s report of fifty thousand such applications.¹⁹ The line to keep from Arnaud’s page is the division of labour: “Arnaud prompted, Emergent built, and the team tested.

The product manager. Take what the owner said in a meeting, turn it into what to build first and what to leave, write it down so the engineer — human or not — cannot misread it. That is a specification, and writing one is now a job the software does well enough that the human’s role shifts to reading it and saying no. Our own specifications are written this way and the arguments over them are in the record.¹ This is the one desk on the list for which we could not find a small firm willing to say in public that software writes its specifications. So the only case here is ours, and it is labelled as such: a company’s claim about itself, checkable on its own record page and nowhere else.

And the one that stays. Somebody reads what went out before it goes out. Not everything — the first fifty, then a sample — but somebody, with a name, whose job it is. This desk does not get handed over, and a firm that hands it over has not automated anything; it has stopped looking. Eighty-nine per cent of small-business clients say they want to know where AI was used in their work.⁵ The checker is how you can answer.

The two things between you and any of this

If the desks above are real, and they are, why is your firm still asking the box questions? Two reasons, and neither is that you are behind.

Nobody set it up. Setting it up is not installing an app. It is deciding, for your firm, which desk first; what that desk may read and may not; what it must do when it is unsure; who checks, and how often. That is a week of someone’s attention who knows both the software and your trade, and it is the week most firms never get, because the vendors sell the software and leave the fitting to you. The previous entry called this the fitting, and argued that it cannot arrive in a box.⁶ It still cannot. The number that says how rare that week is: among Canadian firms of one to four people that use AI, 10.7 per cent used an outside consultant or vendor and 8.2 per cent hired anyone with AI skills, against 30 and 33 per cent at firms of a hundred or more.²⁰ The smallest firms are the ones doing it alone.

Nobody told you what it must never see. This is the one owners are right to worry about and wrong about how. The fear is that “AI” means your client list on someone else’s computer. It does, if you paste it into the box. It does not have to, if the desk is set up with a wall: the software gets the structured facts it needs for the job and is never shown the records that carry your licence — client names where they are not needed, health and legal and financial records, your prices and costs.

Which records go behind the wall is the first thing written down, before anything is installed, and it is written in your words: the chase may see the amount and the date; it may not see the file. A rule you can read is a rule you can check.

For a practice this is not a preference; it is a rule with a regulator behind it. The Canadian Medical Protective Association tells physicians to get the patient’s consent before any AI transcribes an encounter, and to review every note before it enters the record, because the software can invent.²¹ In Alberta a clinic cannot switch an AI scribe on at all until a privacy impact assessment has gone to the Commissioner.²² New York’s bar says the same for a lawyer recording a client call.²³ And the notetaker sitting in on your own calls: Otter is facing a class action alleging it recorded meetings without asking the other people in them, and shared the recordings to improve its systems.²⁴ The wall is not paranoia. It is the licence.

What stays yours

Three things, and they are not small.

Your name. Whatever wrote the first draft, the signature is yours, and the person who receives it will hold you to it. That is why the checker’s desk stays.

The rules. A page, not a binder: what each desk may do, what it may see, who looks before it goes out. Written so your staff can follow it and your insurer can read it.

The record. What it did, when, on what basis. When a client asks — and eighty-nine per cent say they will — the answer is a lookup, not a scramble.⁵

A test you can run this week

Pick the desk that rings at 6pm. Write one page: what it may do, what it may never see, who checks. Run the first fifty through a person. Count how many they changed.

If the answer is most of them, the desk is not ready and you have learned that for the price of a week. If the answer is a handful, you have just done the fitting, and the software can take the desk on Monday with the same person checking a sample.

Either way you now know something the box would never have told you: which of your jobs it can hold.

What we do

We put AI that does the work into a small firm and keep it answerable to the person whose name is on the door. An accountability check in the first week — which desk, what it may see, who checks. A written policy, a page long. Software where each piece does one named job under that policy. And a monthly engagement that keeps the rules current when your business changes, because it will.

The full four steps, and everything we have built this way, are on the services and record pages. The first conversation is diagnostic, not a pitch.