The fact that changes the arithmetic
A manufacturer that takes four hours out of building a unit keeps the money. An accounting firm that takes four hours out of a return does not, at least not automatically, because those four hours were the product. Whether the saving lands with the firm or with the client is decided by the fee arrangement, not by the software. This is the difference between professional services and every other industry that buys automation, and almost nothing written about AI for services firms acknowledges it. The consequence is practical: some hours are worth cutting and some are worth protecting, and knowing which is which is most of the decision.
Every hour a fee earner works falls into one of three pools, and the three behave completely differently when you make them faster.
Hours billed and collected. The client pays the stated rate for them. Make these faster under an hourly arrangement and revenue falls unless the freed capacity gets sold to someone else. Under a fixed fee, the saving is yours.
Hours billed and written down. The work was done, the time was recorded, and the bill went out lower than the time on the ledger because a partner decided the client would not wear it. This pool is pure loss today. Anything that shrinks it converts directly to margin and nobody has to renegotiate anything.
Hours never billed at all. Proposals and pitches, conflict checks, engagement setup, internal reporting, staffing the schedule, chasing timesheets, writing invoice narratives, onboarding paperwork, the search for the memo somebody wrote three years ago. Overhead by definition. Cutting it costs no client a dollar and costs the firm nothing in fee.
The order of attack follows from that, and it is the reverse of the order most firms try. Start with pool three, move to pool two, touch pool one last and only with the fee model decided first.
You are probably here because
- A partner came back from a conference wanting a firm-wide AI plan by the end of the quarter
- Two associates are already pasting client material into a consumer chatbot and nobody has told them whether that is allowed
- A vendor quoted a six-figure platform and you cannot tell what it does that a subscription does not
- Clients have started asking, in writing, what you do with their material
The confidentiality section answers the second and fourth. The cost section answers the third. The first is answered by picking one workflow, not by writing a strategy.
Start with the hours nobody invoices
Take a firm with a hundred and twenty fee earners. Ask each of them how much of a normal week goes to work that never appears on an invoice and is not client work: assembling a pitch from three old ones, hunting for a precedent, reformatting a deliverable, writing time narratives on Friday for a week they half remember. The honest answer in most firms is somewhere between four and eight hours. Take five as the middle. That is six hundred hours a week the firm is paying for and nobody is buying.
At a loaded cost of ninety-five dollars an hour, six hundred hours is roughly fifty-seven thousand dollars a week of capacity, and about three million a year. You will not get it back. Recovering a fifth of it would be a strong result and a fifth is still six hundred thousand dollars of capacity a year, which is a real number to size a project against. It is also the number to hold a vendor to, because a proposal that implies you recover all of it is describing a firm with no meetings in it.
Two cautions on that arithmetic. Freed capacity only converts to money if there is work waiting for it; in a firm turning away engagements it converts almost immediately, and in a firm with soft demand it converts to earlier evenings, which is worth something but is not revenue. And the loaded cost, not the billing rate, is the right multiplier, because nobody was paying the billing rate for these hours in the first place.
Where the payback is, as we weight it when scoping a firm
Editorial weighting from our own scoping practice, shown to explain how we rank the options. Not a measured statistic.
The bottom row is the one firms ask for first and it is last on the list for a reason. A public assistant is the highest-risk, lowest-return thing a services firm can build: strangers on the internet, unbounded questions, an answer that reads as advice, and no fee attached to any of it.
Confidentiality is the gate, and it is a contract problem before it is a technology problem
The blocking question in a services firm is almost never "can the model do this." It is "may this material leave the building, and who agreed to that." Four documents decide it, and the firm usually has all four already.
Your professional standards. For accountants, the AICPA Code of Professional Conduct at 1.700.001 forbids disclosing confidential client information without the client's specific consent. In tax practice the stakes rise: Internal Revenue Code section 7216 makes a preparer's knowing or reckless disclosure or use of tax return information a misdemeanor, and section 6713 attaches a civil penalty on top. For lawyers, ABA Formal Opinion 512, issued in July 2024, is the first national ethics guidance on generative AI and treats confidentiality, competence, client communication and fee reasonableness as live duties rather than aspirations. Read the one that governs your firm before the vendor call, not after.
Your engagement letters. Some permit subcontractors and cloud processing; some are silent; some, particularly with large corporate clients, contain outside counsel guidelines or supplier terms that name specific technologies. Silence is not permission, and it is also not prohibition. It is a question somebody has to answer per client.
Client-imposed terms. Financial and healthcare clients increasingly send terms that forbid processing their material with a third-party model, or require notice, or require that nothing be retained for training. These are usually satisfiable. They are rarely satisfiable retroactively, after a system is already indexing everything.
Your own data-protection posture. Which vendors already hold this material lawfully, what your retention schedule says, and what your insurer was told.
The practical output of that reading is a small table: client or matter type, what the terms permit, and which of the firm's tools each category may use. Two weeks of somebody careful. It is the least interesting artifact in the project and it is the one that stops the project being unwound in month six.
The four things firms ask for, and what each really costs
Across the requests that arrive from services firms, four cover most of the ground. They differ far more in effort than they do in description.
| What is asked for | What it really requires | Where it goes wrong | Honest range |
|---|---|---|---|
| Draft this from our precedent Proposals, scopes, standard memos | A curated set of good examples, not the whole drive; a review step; a house style captured explicitly | Training it on everything, including the bad ones, then wondering why the drafts are mediocre | $45K–$110K |
| Read what the client sent us Contracts, statements, drawings, records | A labeled test set with correct answers, a citation back to the source page, and a measured error rate | Buying a demo on ten clean documents, then meeting the forty layouts nobody mentioned | $70K–$180K |
| Find what we already know Search across prior work product | Permissions carried into the index, matter-level access control, and a way to exclude what must stay walled | An index that ignores ethical walls and shows an associate a matter they may not see | $90K–$220K |
| Write our time narratives From calendar, documents and system activity | Access to the systems that already log activity, plus a fee earner who confirms before it posts | Posting entries without confirmation, which turns a billing question into a professional one | $35K–$90K |
Those ranges are engineering cost for a first production version at a firm of a hundred to five hundred people, and they exclude your model and software subscriptions. Plan on a recurring cost afterwards of roughly fifteen to twenty-five percent of the build price a year, covering hosting, model usage, monitoring and the changes that follow the first six months of real use.
Knowledge reuse compounds, and is the hardest of the four
Every firm over about fifty people has the same asset and the same problem: twenty years of work product that is worth a great deal and cannot be found. A partner knows the firm did something similar in 2019 and cannot remember for whom. The associate rebuilds it from scratch in eleven hours. The eleven hours get written down.
Search over that corpus is the highest-value thing on the list and the one most likely to be built wrong, because the requirement that matters is not retrieval quality. It is access control. The document management system already knows who may see which matter. The moment you copy that content into a search index, you have created a second system that does not know, unless somebody built it to. Ethical walls, client-imposed restrictions and departed-employee matters all live in the first system and none of them travel by default.
This is the requirement that gets cut when the budget tightens, because it is invisible in a demo. Everything works in a demo; the demo is run by a partner who can see everything. It fails in month four when an associate on a screened matter gets a result they should never have seen, and the firm now has a reportable problem rather than a software problem.
The index must enforce the same access rules as the source, per user, at query time
Ask any vendor exactly how permissions reach the index, how quickly a change to a wall propagates, and what happens to a document whose access changes after indexing. A good answer is specific and slightly tedious. A bad answer is that everyone in the firm can already see everything, which is true in very few firms and never true in the ones with the most valuable corpus.
What it costs, and the cheapest honest first step
A single workflow taken to production, used daily by real people, generally lands between sixty and a hundred and eighty thousand dollars of engineering and six to fourteen weeks of calendar time. Tightly bounded work comes in under that. Anything touching regulated material, or replacing something a hundred people already depend on, goes above it.
Before any of that, the cheapest honest step is a short scoping block that ends in a written scope, an acceptance test with a number in it, and a data-access assessment naming every blocker and who owns it. Priced separately, that artifact belongs to the firm and can be handed to any competent shop, including one that is not the shop that wrote it. Bundling scoping into a build makes you captive to whoever ran it, which is a bad thing to be even when the vendor is good.
Send the workflow and we will tell you what it would take.
Describe one workflow in a paragraph, say roughly how many people do it and how often, and note anything in your client terms that constrains it. Email contact@precisionfederal.com. You get back a short written note: whether we would build it, what we would build first, and the range. One business day, no charge, no meeting.
contact@precisionfederal.comThe pricing conversation the partners have to have
Suppose a piece of recurring work gets thirty percent faster. Three things can happen to that thirty percent, and the firm chooses which.
The client keeps it. Hourly billing, same rate, fewer hours on the invoice. Revenue on that engagement falls by thirty percent. This is the default outcome if nobody decides anything, and it is why some firms quietly discourage the tooling they just bought.
The firm keeps it. Fixed fee agreed before the efficiency existed. Margin rises. This is the honest case for moving a standardized service to a fixed price, and it works best on work the firm does many times where the variance is already low.
Nobody keeps it and the capacity is resold. Hourly billing, same hours worked, more matters through the same people. This is the best outcome and it requires demand to exist. It is a sales question wearing a technology costume.
There is a fourth consequence that is not financial and gets raised late. Junior work is how juniors learn. A firm that automates the first-pass review has removed the exercise that turns a second-year into a fifth-year, and it will notice in about four years. That is not an argument against doing it. It is an argument for deciding, deliberately, what replaces the training that the task used to provide.
How this fails in practice
- A firm-wide platform bought before a single workflow was measured, so nobody can say whether it worked
- An index built over the whole document system with no matter-level permissions, discovered by an incident
- Drafts trained on every past document, good and bad, producing output at the firm’s average rather than its best
- No acceptance test, so the argument about whether it is working is settled by whoever spoke last
- A pilot with the three enthusiastic partners, whose result predicts nothing about the other forty
- Client terms read after the build, forcing a rebuild or a set of awkward letters
- Time narratives posted without a human confirming them, turning a convenience into a billing exposure
- Efficiency delivered under hourly fees with no fee-model decision, so the firm funded a client discount
A first ninety days that produces something
First engagement, one workflow
Step five is the one that is skipped and the one that pays. Running in parallel on live work is the only way to learn what the tool does on a bad day rather than a chosen day, and it is the only comparison a skeptical partner will accept.
Before you fund it
- One named workflow with a measured current cost in hours
- A written acceptance test with a number in it, agreed before the build
- A permission table covering every client whose material will be touched
- A named owner for each data source and a date access will exist
- Access control specified for the index, not assumed from the source system
- A decision on who keeps the saved time, made by the partners in advance
- A run-cost estimate for year two, not only a build price
- A named person inside the firm who will own it after handover
Bottom line
Firms that sell hours should automate the hours nobody is buying first, because that saving is unambiguous and requires no negotiation with anyone. The gate on everything else is contractual: what the professional standards, the engagement letters and the client terms permit, which is knowable in a fortnight and expensive to discover late. The most valuable target is the firm's own accumulated work product, and it is the one where access control decides whether the project is an asset or an incident. Pick one workflow, measure it, test it against real examples, and settle the fee question before the efficiency arrives rather than after.
Frequently asked questions
Buy where the work is generic and the vendor already has your data in their system: document management search, transcription, calendar and time capture. Build where the value comes from your own material, your own house style, or a workflow that is genuinely how your firm works. The costly mistake is buying a general platform for a specific problem, then paying twice over in configuration to make it fit.
It depends on your profession and the engagement terms, and it is a question for your general counsel or your risk partner rather than your vendor. Accountants are governed by the AICPA confidentiality rule and, in tax work, by Internal Revenue Code sections 7216 and 6713. Lawyers should read ABA Formal Opinion 512 alongside their state rules. Corporate clients increasingly impose their own terms that are stricter than either. Build the permission table before the system, not after.
For one workflow taken to production and used daily, plan on roughly $60,000 to $180,000 of engineering and six to fourteen weeks, plus recurring cost of about fifteen to twenty-five percent of the build price per year. A short scoping block ahead of that, priced separately, gives you a written scope and an acceptance test you can take to any firm.
On hourly engagements, yes, unless one of two things happens: the freed capacity gets sold to another client, or the service moves to a fixed fee agreed before the efficiency existed. That is why the first projects should target unbilled overhead and written-down time, where the saving is the firm's by default and no fee conversation is required.
A held-out set of real examples with correct answers attached, scored by a script rather than by impression. Two hundred to five hundred examples adjudicated by someone who knows the work will settle nearly every argument you are going to have, including the one where a stakeholder finds a single bad output and concludes the whole thing is unreliable. One example is an anecdote. A scored set is evidence.