Roles covered: proposal manager, bid coordinator, proposal writer, content-library owner. Functions: sales, business development. Linked workflows: proposals, project knowledge retrieval, meeting follow-up.
Your most important output is a response the organization can stand behind. Faster writing helps only when the proposal meets the request, uses permitted evidence, and makes promises delivery teams have approved.
This chapter helps you decide what AI should and should not touch in your proposal process, who must approve what, and how to run a first quarter without trading accuracy for speed. The worked case, economics, and evaluation design for the underlying workflow are in the proposal chapter; this guide does not repeat them. The planning example below uses an invented firm. Recommended practices are proposals for local agreement, not findings from a deployment.
The decisions you own
Three decisions sit with the proposal function rather than with engineering, finance, or a vendor.
What counts as a finished response. You define the standard a draft must meet before partners see it: every mandatory requirement answered or explicitly unresolved, every material claim traceable to permitted evidence, and every commitment matched to an approved scope. A tool that produces more pages against a lower standard has made the process worse.
Which sections may be drafted and which must be authored. You propose, and partners ratify, a boundary between content that assistance may prepare and content that an accountable person must write or approve line by line. Scope, price, staffing, delivery dates, and contractual commitments belong in the second category. So does any statement about a named client. In this guide's proposed model, contractual commitments are approved by an authorized partner, with legal review where the firm requires it.
Where assistance may write, and where it may not
- May be drafted, then checked
- Authored or approved line by line
- Unchanged or ordinary software
Proposal outline
1. Response to requirementsMay be drafted
Proposal manager confirms every mandatory requirement is answered
2. ApproachMay be drafted
Delivery lead confirms feasibility
3. Credentials and case studiesDrafted only from permitted evidence
Evidence owner approves reuse; any statement about a named client is authored or approved
4. Team and staffingAuthored or approved line by line
Delivery lead proposes named staff; authorized partner approves; no drafting at first
5. Scope and delivery datesAuthored or approved line by line
Delivery lead estimates; finance and partner approve scope against price
6. Price and commercial termsAuthored or approved line by line
Finance and authorized partner approve
7. Contractual commitmentsAuthored or approved line by line
Authorized partner, with legal review where the firm requires it
8. Formatting and assemblyOrdinary software and document control
Proposal team, using existing tools
How the evidence library is governed. Someone must own each reusable credential, case summary, biography, and method description: its date, permitted audience, and retirement. If you do not assign that ownership, a retrieval system will reuse whatever it can find.
You influence, but do not own, the bid decision. Partners decide whether to pursue an opportunity. Your contribution is making the decision explicit and the evidence for it visible. A drafting tool that makes bidding cheaper can quietly turn every inbound request into a bid; guarding against that is part of your role.
The measure that looks like progress but is not: proposals produced per month, or first-draft turnaround. Both can improve while missed requirements, unsupported claims, and partner rework rise.
Map the work you manage
The proposal process contains several distinct tasks with different consequences when they go wrong. Start with the tasks where information and authority are clear.
Capability verbs in the table follow the capabilities chapter; starting authority uses the arrangements in assistance, approval, and autonomy.
| Task | Consequence of an error | Candidate capability | Proposed starting authority | Canonical chapter or first fix |
|---|---|---|---|---|
| Qualification summary | Pursuing a poor-fit bid, or missing a disqualifying condition | Extract, summarize | Assistance; partner decides | Proposals |
| Requirements and compliance matrix | Missed mandatory requirement; disqualification | Extract, classify | Assistance; proposal manager confirms coverage | Proposals, document extraction |
| Finding prior work and people | Reusing restricted client material; wrong expert | Retrieve | Assistance; permission-filtered | Project knowledge |
| Discovery-call notes into requirements | A buyer's tentative comment treated as a requirement | Extract, summarize | Assistance; call lead confirms | Meeting follow-up |
| Credentials and case studies | Unsupported or unpermitted claim reaches a buyer | Retrieve, generate | Approval before action; evidence owner signs off | Proposals |
| Approach narrative | Promises the delivery team cannot keep | Generate | Approval before action; delivery lead confirms feasibility | Proposals |
| Staffing and biographies | Named person unavailable or misdescribed | Retrieve | Unchanged at first; staffing owner confirms | — |
| Pricing and commercial terms | Unprofitable or unpriced scope | None in the pilot | Unchanged; finance and partner decide | Economics |
| Formatting and assembly | Wrong version submitted | Ordinary software | Unchanged; use existing tools | — |
Two rows deserve a simpler fix before any model. If submissions go out in the wrong version, the problem is document control. If requirement matrices are rebuilt from scratch for each bid, a maintained template may capture much of the gain. Compare against those alternatives before attributing improvement to generation.
The work you perform (matrices, library curation, coordination) differs from the work you supervise (writers drafting sections, subject-matter experts contributing). Assistance changes both, but the second requires you to redesign review, not just adopt a tool.
Who decides what
A consequential proposal error need not be a writing error. It can be a decision made by the wrong person, or not made at all.
| Decision or activity | Proposal manager | Writers | Evidence / knowledge owner | Delivery lead | Finance | Authorized partner | Engineering / IT |
|---|---|---|---|---|---|---|---|
| Pursue or decline | Prepares brief | — | — | Advises on capacity | Advises on economics | Decides | — |
| Requirement coverage | Owns | Contribute | — | Confirms technical items | — | Reviews | — |
| Reuse of credentials and client names | Requests | — | Approves | — | — | Escalation | Enforces permissions |
| Approach and milestones | Coordinates | Draft | — | Approves feasibility | — | Reviews | — |
| Named staff | Requests | — | — | Proposes | — | Approves | — |
| Price and scope match | Flags mismatches | — | — | Estimates | Approves | Approves | — |
| Contractual commitments | Flags terms in the request | — | — | — | — | Approves, with legal review where the firm requires it | — |
| Final submission | Assembles | — | — | — | — | Submits | — |
| Assistant configuration and access | Specifies needs | Report failures | Defines access classes | — | — | — | Operates |
Owning the outcome, operating the system, and approving an external action are separate responsibilities. Engineering can make sure the assistant cannot retrieve a restricted engagement file; it cannot decide whether a credential may be quoted. You can flag that a draft's scope exceeds the estimate; you cannot approve the price.
Plan a first quarter
The following planning example is fictional. Fallowgate Advisory is an invented operations consultancy with a three-person proposal desk supporting about a dozen partners. The numbers and decisions exist to show the sequence, not to describe any real firm.
Weeks 1–3: measure before choosing a tool. The proposal manager records, for recent opportunities, time by stage and contributor, time spent waiting for decisions, missed or late requirements found at partner review, and bids later judged not worth pursuing. The finding that matters in this invented case: requirement extraction and matrix-building consume a large share of desk effort on public-sector and procurement-led bids, while partner delay on pricing controls the deadline for private-sector bids. Reducing drafting time would not shorten the second kind.
Weeks 3–4: choose a first scope. The desk selects requirement extraction and compliance matrices for procurement-led bids. The reasons: the input is a document, the output can be checked against that document, errors are detectable before submission, and no client-confidential material is needed. The proposal manager writes down what the pilot will not do: no pricing text, no named-client case studies, no staffing statements.
Weeks 5–8: offline test. Using past requests with known final matrices, the desk tests whether proposed matrices capture every mandatory requirement, handle addenda that change deadlines or requirements, and mark unclear requirements as questions rather than guessing. The acceptance rule, agreed before testing: no mandatory requirement silently omitted in the test set. Setting that rule, rather than an average accuracy, is the proposal manager's job. The pilot design playbook covers sample construction; the proposal chapter lists the test variations.
Weeks 9–12: supervised use and decision. Writers use the assisted matrix on live procurement bids while still checking each requirement against the source. The desk records correction effort, requirements missed and caught, and partner review time. At the end of the quarter, the desk and the sponsoring partner decide to expand, revise, or stop, using the criteria set in week 4.
The counterexample: what Fallowgate should not start with. A partner asks for generated case studies and pricing narratives because those sections are slow to write. Both are poor first choices. Case studies name clients and cite results, so each one requires permission and evidence checks that a drafting tool does not remove; the knowledge-retrieval chapter shows how restricted engagement material can surface through retrieval. Pricing narratives describe commitments whose correctness depends on a commercial decision the tool cannot make. Starting there would create the most serious possible errors in the least checkable sections. The right response is to fix the evidence library first and keep pricing text authored.
Measure the right things
| Measure | Why it matters | How it can mislead |
|---|---|---|
| Mandatory requirements missed at partner review or after submission | The failure buyers penalize most directly | Requirements caught late by partners look like success if only submitted proposals are counted |
| Unsupported or unpermitted claims per draft, caught and uncaught | Tests whether evidence controls work | A strong reviewer can hide a weak assistant; record errors in drafts, not only in submissions |
| Total labor per qualified opportunity, by role | Captures review and rework | Shifting effort from the desk to partners can look like savings on the desk's timesheet |
| Elapsed time from qualification to approved submission | Deadline pressure is real | Controlled by decision delays as much as drafting |
| Bid/no-bid decisions recorded, including declines | Guards against over-bidding | Win rate can rise simply because fewer marginal bids were submitted |
| Library maintenance effort and overdue items | The continuing cost you must fund | May sit outside the proposal budget and go unmeasured |
| Win rate and deal size | The eventual business question | Bid selection, relationships, and pricing can overwhelm any drafting effect |
Report win rate last and cautiously. A pilot can establish effort and quality changes well before it can establish a sales effect.
Your team: work, judgment, and skills
What may shift. If the pilot works, writers may spend less time on first-pass extraction and more on checking coverage, resolving questions with delivery leads, and tightening the argument. Whether that shift actually occurs is something to measure in your pilot, not assume. The proposal manager's time may move toward library governance and coordination.
What judgment remains. Deciding what the buyer is really asking for, what the firm can credibly claim, and how to structure a win strategy. Recognizing when fluent text smooths over an unresolved question. Knowing when to recommend declining.
Skills to build, by group:
- Proposal manager: designing evidence standards and acceptance rules; running a small evaluation; negotiating section boundaries with partners.
- Writers: reading a requirement against its source; building and using an evidence matrix; identifying unsupported claims and quiet commitments in generated prose.
- Evidence and library owners: labeling permitted use and audience; retiring stale credentials; approving sanitized case summaries.
- Partners and delivery leads: reviewing commitments rather than prose; recording decisions so drafts can reflect them.
Protect learning. Junior writers learn to build a proposal by doing it: reading requests closely, deciding structure, arguing for an approach. If assistance removes that practice, the desk may lose the ability to handle unusual requests that fall outside the library. A proposed safeguard: juniors author some responses from first principles each quarter, and every reviewer explains substantive corrections so they become teaching material.
Treat any change to headcount, role definitions, or career paths as a decision to plan and measure, not as a forecast. This guide does not report observed changes in proposal teams.
Who to bring in. A first proposal pilot can bring in outside help as a builder for the requirement-extraction tool, as integration work to connect it to the bid library under the access classes the evidence owner defines, and as an independent evaluator to build the test set of past requests and addenda and check results against the no-silent-omission rule. Keep in-house what the "Who decides what" table assigns to the firm: requirement coverage, approval of credentials and client names, price, named staff, and submission. An outsider can test a matrix; only the firm can decide what it may claim. See Staffing an AI pilot and Choosing and working with an AI partner.
Change management and governance
Involve the writers early. They know where requests hide mandatory conditions and which library items are unreliable. Ask them to define the failure categories before the pilot.
Change the targets. If writers are measured on pages produced or turnaround alone, they are rewarded for accepting unsupported drafts. Add requirement coverage and caught claims to how work is reviewed.
Make reporting failures cheap. A writer who finds an invented credential should be able to flag it in one step, and the report should reach the library owner and the pilot lead. Treat caught failures as evidence, not embarrassment.
Set the rule on client material in writing. Which past proposals and deliverables may be retrieved, by whom, and for what purpose. Contractual confidentiality and conflict arrangements vary by firm and client; the rule should be agreed with whoever owns those obligations.
Guard against automation bias. The authority chapter describes the risk and the signs that review has become a formality. The proposed response is structural: reviewers check claims against sources, not drafts against their sense of plausibility.
Warning signs
| Signal | What it may indicate | Proposed response |
|---|---|---|
| Bid volume rises while qualification notes get shorter | Cheaper drafting is displacing the bid decision | Require a written pursue/decline rationale before drafting |
| Partners report more rework, desk reports less effort | Work has shifted upward, not disappeared | Measure labor by role across the whole cycle |
| Drafts cite credentials no one can locate | Library lacks ownership or the assistant is filling gaps | Pause affected sections; audit the library |
| Writers stop opening the source request | Over-reliance on the matrix | Spot-check coverage against the original; reinforce the review standard |
| Past client names appear in drafts for unrelated bids | Permission filtering is missing or wrong | Treat as a serious failure; stop and investigate retrieval access |
| Library items older than their review date | Maintenance is unfunded | Assign owners and time before expanding |
These are things to watch, not reported frequencies.
Questions to ask vendors and internal teams
- How does the system enforce which documents a user may retrieve, before retrieval rather than after drafting?
- Can every material sentence in a draft be traced to a specific source passage and version?
- How does it handle an addendum that supersedes part of the original request?
- What happens when a requirement has no supporting evidence: does it leave a visible gap or write around it?
- Can we export reviewer corrections and use them to evaluate new versions without training on client material?
- What does it cost per opportunity, including retries and the time our reviewers spend?
- Can we turn it off for particular sections, clients, or bid types?
- Who at our firm approves adding a document to the retrievable library?
When to wait
Defer if the evidence library has no owners, if reuse permissions for past client work are unclear, or if most proposal effort goes into bespoke scoping and commercial negotiation rather than assembly. Defer if partners are not willing to specify which sections require their authorship. In those conditions, the better first investments are a qualification checklist, a maintained requirement-matrix template, and a curated credential library with owners and dates. Those improvements help whether or not generated drafting follows.
What external research establishes
For your analyst
Dell'Acqua and coauthors report a randomized experiment with 758 consultants in which AI assistance had task-dependent effects. On a task designed to fall outside the model's capabilities, correctness was 84.5% without AI and 60% and 70.6% in the two AI conditions, about 19 percentage points lower when those conditions are combined. The model tested dates to June 2023. Organization Science, 2026 We read the main paper; we did not verify its supplementary appendix or preregistration.
The study does not measure proposal work, current models, or firm outcomes. The proposed implication for this role is narrow: judge commitments, evidence, and prose separately, and include cases where the right answer is to leave a gap or decline, because a fluent draft can be wrong on the point that matters.
Google's engineering guidance recommends defining metrics and starting with a simple system before adding machine learning. Rules of Machine Learning Applying that principle to proposals, by comparing against a maintained template first, is an editorial recommendation.
No attributable evidence was found on how AI assistance changes proposal-team effort, win rates, or staffing. This guide makes no such claim.
Related reading
Proposal workflow, project knowledge retrieval, meeting follow-up, professional services, human review, and finance leaders. Prepare a pilot with the pilot kit and evaluation scorecard.