Professional services firms do not lose first on the signature. They lose on the factory that feeds it.
A two-person shop with an agent stack cannot take the partner’s name, the license, or the relationship. It can take the repeatable work that turns a client file into a first draft the partner still has to own: the diligence memo, the return, the issues list, the workpaper set, the redline.
That factory is the line the 70% question is asking about. The first two workflows are the two acts inside it that can produce a comparison a managing partner will read.
This is the same selection method as the note on choosing without politics. The tests do not change. The costumes do. Partnerships wrap every decision in independence, privilege, professional skepticism, and the billable hour. Those are real constraints. They are also how a firm picks a chatbot and calls it transformation.
A workflow is still a sequence with an input, a committed output, a gate, and a system of record. “Audit,” “tax,” “corporate,” and “deal advisory” are practices. They are not workflows.
Attractors in a partnership
Safety has a local shape here.
The photographable assistant. A bot that answers the intranet FAQ. Copilot that drafts the holding email. A “matter companion” shown at the partners’ offsite. Nothing is filed. Nothing is deprecated. Liability is untouched. The photograph is the product.
The unowned archive. Tagging twenty years of opinions. Translating standard engagement letters. Cleaning the know-how database no one searches. Easy access, no client, no deadline, no moving number.
The craft that is also the identity. The audit partner’s review notes. The tax partner’s “I know this inspector.” The corporate partner’s markup style. The argument becomes whether machines are allowed to be skeptical, not whether a parallel file can beat cycle time and error rate on the same pack. You never reach the run.
The platform before the file. “Ingest the DMS, build a knowledge graph of every matter, fine-tune on our opinions, then we will see.” That is a research program with a data-protection appendix. The twin has to start on the pack that arrived this morning.
Independence rules, privilege, and filing deadlines are not attractors. They belong in the gate and the rollback. Using them to postpone naming a workflow is the attractor.
Practices are not candidates.
These fail before the tests:
- “AI in the audit practice.”
- “A lawyer’s assistant.”
- “Reinvent how we do deals.”
- “Predictive risk scoring across the client base.”
The first three have no input and no output. The fourth is a model looking for a process. Write from this pack to that filed object or start again.
Throughput in this industry is not the three bet-the-firm mandates of the year. It is the file that already repeats: recurring statutory audits for a book of similar clients, compliance returns, incoming third-party paper, mid-market sell-side or buy-side packs, standard SPA comment matrices. If a four-week parallel cannot show tens of instances, you chose the mandate that makes the brochure.
The compressible core is retrieval, assembly, checking, routing, and drafting against a playbook: pull last year’s file, extract the PBC items, map the data-room index, compare the clause to the standard position, list the open points, build the first memo. The gate is narrow — opinion, materiality, departure from the playbook, independence, whether the letter goes out.
If most of the clock is a senior person hunting and pasting, you have a workflow. If most of the clock is a partner inventing the theory of the case, you have a craft. Do not open there.
Live packs only. Same client, same data room, same inbox. Outputs that already exist in the DMS, the tax engine, or the deal workspace: draft return, first workpaper set, issues list, redline, diligence memo, Q&A log.
Baselines that already live in a time-and-billing extract even if nobody loves them: days to first draft, review rounds, first-pass completeness, leakages found after filing, hours per file, not hours “saved.”
Blast radius is the serious test in this sector. A draft that never leaves the review queue is reversible. A filed return, an audit opinion, a sent markup that waives a point, a memo that breaks privilege — those are not the first twin. Design for “held for the named reviewer.” Do not design for “already on the portal.”
Ownership cannot be “the partnership.” Name the practice lead who can change the path for that file type without a vote: head of tax compliance, TAS / transaction-tax lead, audit methodology for a recurring book, head of contracting for a product line, deal-advisory operating partner. If the first action is a partners’ meeting, you picked the identity, not the factory.
Two pairs that survive
Pick the pair that sits on the line the 70% question exposed. Do not pick one from audit and one from M&A to keep two department heads quiet. Two workflows exist, so one stack can generalize across adjoining files.
Pair A — when the exposed line is the transaction factory
This is usually the higher-margin copyable line: mid-market M&A, transaction tax, legal diligence feeding a sale or acquisition.
Workflow 1. Complete data room/facts pack → first-draft diligence findings memo.
Input: an index plus the documents that are actually in the room — contracts, accounts, tax filings, employment paper, the management Q&A.
Output: a structured findings memo in the deal workspace: issues ranked, source tagged, playbook position noted, open questions listed.
Gate: the deal lead or legal/tax diligence lead releases the draft to the client or the SPA team.
Why it is a workflow: the room repeats; the work is extract–compare–assemble; cycle time and missed issues are already fought over; a bad draft stays inside; one named lead can own the path.
Workflow 2. Released findings memo → first-draft issues schedule for the SPA/tax structure/disclosure letter.
Input: the memo from workflow 1 plus the current draft agreement or structure paper.
Output: a comment matrix or issues schedule tied to clause or step, with recommended position and fallback.
Gate: the partner on the instrument.
Why it adjoins: it consumes the structured findings: same room, same playbook, same issue taxonomy. The stack does not start again at zero when the SPA arrives.
This is pack-to-memo-to-instrument. It is not “AI for deals.”
Pair B — when the exposed line is the recurring file
Compliance and statutory work is less glamorous and often more copyable. Volume is the point.
Workflow 1. Client source pack → draft return or draft workpaper set.
Input: the standing PBC or source bundle — trial balance, payroll extracts, bank files, contracts flagged as recurring, last year’s file.
Output: a draft in the tax engine or the audit binder, with every number and position traced to a source and every gap on an exception list.
Gate: the manager. The partner is not the preparer.
Why it is a workflow: it happens on a calendar; most elapsed time is gathering and tying; error rate and days-to-draft already exist; nothing is filed without the gate; ownership sits with compliance or audit operations, not with the whole partnership.
Workflow 2. Review notes + client queries → cleared file ready for partner sign-off.
Input: the draft from workflow 1 plus the review comments and the client answers that always arrive late.
Output: a cleared binder or return pack with a short exception log of what remains for the partner.
Gate: the partner signs. That gate stays human.
Why it adjoins: same file, same traces, same engine; the learning loop is whether the twin’s first draft reduced review rounds, not whether it replaced the signature.
Do not open with opinion work, courtroom advocacy, or the one-off cross-border restructuring that will be discussed for nine months. Those fail throughput and the craft test.
What the stack does to a file
The five layers have to operate on the pack you already received.
Sense. Watch the data room, client portal, intake mailbox, and PBC folder. Ingest PDFs, workbooks, markups, chat exports. Do not wait for someone to forward the zip.
Interpret. Structure against your taxonomy: clause type, account assertion, tax position, document family. Retrieve last year’s file, the similar deal, the playbook position. Extract the facts that later become exceptions — change-of-control, related parties, non-standard indemnities, missing PBC items.
Decide. Propose the draft under the playbook. Flag the minority that leaves the playbook. Do not invent a new opinion. Do not silently take a position that used to need a partner’s initials.
Orchestrate. Write the memo, the matrix, the draft return, the workpaper set into the systems of record you already use — DMS, deal workspace, tax engine, audit binder. Route only the exceptions. Do not create a sixth place where the file lives.
Learn. Compare the twin’s draft with what the manager released and what the partner signed. Cycle time, first-pass completeness, issues found after release, cost per file, incidents (wrong number, missed clause, privilege leak). Two workflows exist so you can see whether the loop transfers from memo to instrument, or from draft to cleared file.
Humans stay on the gate. They are above the loop.
Same pack, two paths
The official file continues to be prepared the old way. The twin takes the same pack, writes into a review queue, and leaves a trace: which prior file it used, which playbook rule fired, which field it could not fill.
Once a week, the named owner and the managing partner read four numbers per workflow: time to draft, quality (completeness or error rate), cost per file, risk events. No utilisation. Not “hours returned to the partner,” which is how leverage politics re-enters. Not a demo at the offsite.
When the twin is better on those four, the old preparer path is retired for the standard majority. The partner gate remains. Then the next adjoining act on the same chain — the Q&A log that feeds the memo, the disclosure letter that follows the issues schedule, the next year’s PBC that follows this year’s cleared file.
Four weeks will tell you if the file actually repeats. Empty queues mean you chose the mandate.
Fill this page
If it takes a workshop, the candidate is not ready.
Workflow 1 — data-room pack to first-draft findings memo
- Input: complete room or facts pack for an in-scope deal type.
- Output: structured memo in the deal workspace, sources tagged, exceptions listed.
- Gate/owner: diligence lead. Releases the draft. Reports to the managing partner for the twin.
- Metrics: days to first memo; issues missed that later appear in the SPA; cost per pack; client questions caused by an incomplete first draft.
- Why the core is coordination: most of the clock is finding the document, comparing it to the playbook, and writing the row.
- Rollback: nothing leaves the queue without the gate. Privilege and circulation rules are part of the queue, not a later policy.
- Reporting: weekly four-number sheet on the same packs the legacy path processed.
Workflow 2 — findings memo to first-draft issues schedule
- Input: released memo plus current draft instrument or structure paper.
- Output: clause-tied comment matrix with position and fallback.
- Gate/owner: partner on the instrument.
- Metrics: days from memo to matrix; comments that survive first negotiation; issues that appear only after signing; cost per matrix.
- Why it adjoins: it is the same taxonomy written one stage downstream.
- Rollback: the matrix is not sent until the gate.
- Reporting: same sheet, same line to the managing partner.
On Pair B, rewrite the page as source pack → draft return or workpapers, then review notes → cleared file. Keep the partner on the last gate.
Refuse these first objects
Do not start with the courtroom or the tax-controversy letter. Judgment-heavy, low frequency, identity-loaded.
Do not start by replacing the DMS. That is the science project.
Do not start with anything that files, signs, or sends without a queue. The twin has not earned that radius — and the insurer will agree.
Do not start with a firm-wide assistant. It will never force a deprecation.
Do not start with the broken file that has been a quality finding for five years. The twin will inherit the fog.
Do not start where a vendor already has a demo trained on generic contracts or generic 10-Ks. Their corpus is not your playbook.
What to do next
Sit with the managing partner and two practice leads. Leave the digitalization committee out of the room.
Write the exposed line in one sentence — transaction factory or recurring file. Put two adjoining workflows on one page. Run the five tests. If either workflow is an assistant, an archive, a craft identity, or a platform program, discard it.
Stand the twin up inside the firewall, on packs you already receive, with humans on the gates and a circulation rule that protects privilege. Run both paths for four weeks. Read the four numbers. Retire the preparer path that loses. Keep the signature. Take the next file on the same chain.
The partnership will keep selling judgment. The question is whether the factory that feeds that judgment still has to be a pyramid of hours.