Free template · AI integration proposal

AI integration proposal template

A proposal for adding an AI feature to an existing product, written so the client knows how often it's right before they rely on it. It starts with an evaluation set, keeps a person in the loop, and is plain about data and cost. Written out in full for a fictional legal-tech company.

Sections
10
Estimate
7 lines, 88 hours
Example total
$12,320.00
Payments
30% / 20% / 30% / 20%
Proposal

Lumen Legal: AI summaries for contract intake

Prepared for Lumen Legal

Investment
From $12,320
Payments
4 stages

What you asked for

Your customers upload around 1,200 contracts a month, and your paralegal team spends twenty minutes on each writing a summary: parties, term, renewal, termination and liability caps.

We propose adding an AI-drafted summary that a paralegal checks and approves, measured against your own contracts before it reaches a customer, so you know how often it's right before you rely on it.

How we'll know it works

  1. An evaluation set first. Your team picks 60 real contracts and writes the summary they'd want; that's the answer key.
  2. A target agreed up front. For example: every party, date and amount correct in at least 95% of summaries, and nothing invented.
  3. Measured on every change. Each version of the prompt or model is scored against the set, so improvements are numbers, not impressions.

If the target can't be met, you'll know after phase 1, before the integration is built.

Scope of work

Phase 1: prove it

  • Evaluation set and scoring script, run on demandS-1
  • Prompt and model choice, compared on cost and accuracyS-2
  • A short report: the score, where it fails, the cost per contractS-3

Phase 2: build it in

  • Summaries drafted on upload, shown to a paralegal to check and approveS-4
  • Every field linked to the sentence it came fromS-5
  • Clear "couldn't find this" instead of a guessS-6

Phase 3: run it

  • Monitoring of cost, speed and approval rateS-7
  • A weekly sample re-scored against the evaluation setS-8
  • A runbook for when the model provider changes somethingS-9

Not included

  • Drafting or redlining contractsX-1
  • Answering legal questionsX-2

Estimate

Phases 1 and 2 at a fixed price. Phase 3 is the set-up; ongoing tuning can be a small monthly retainer.

TaskHoursRateAmount
Phase 1 · 30 h$4,200.00
Evaluation set tooling and scoring script12$140.00$1,680.00
Prompt and model comparison14$140.00$1,960.00
Phase 1 report4$140.00$560.00
Phase 2 · 48 h$6,720.00
Drafting on upload, with source links24$140.00$3,360.00
Review and approve screen for paralegals16$140.00$2,240.00
"Couldn't find this" handling and edge cases8$140.00$1,120.00
Phase 3 · 10 h$1,400.00
Monitoring, weekly re-scoring and runbook10$140.00$1,400.00
Optional extras
Clause library lookup (compare to your standard positions)20$140.00+$2,800.00
Subtotal$12,320.00
Total$12,320.00

Your data

  • Contracts are sent to the model provider under its business terms: not used for training, and kept for the shortest retention it offers
  • Processing stays in the EU region if your customers need it
  • Nothing is sent to an AI model until a customer has agreed to it in your terms
  • The evaluation set stays in your systems

Timeline

About seven weeks: phase 1 in weeks 1–2, phase 2 in weeks 3–6, phase 3 in week 7.

Payment schedule

  1. To startOn approval$3,696.0030%
  2. Phase 1 reportWeek 2$2,464.0020%
  3. Summaries live for reviewWeek 6$3,696.0030%
  4. Monitoring in placeWeek 7$2,464.0020%

Assumptions and running costs

  • At current model prices a summary costs a few cents; phase 1 measures the exact figure on your contracts, and you pay the provider directly
  • Your team writes the 60 reference summaries (about a day for two paralegals)
  • A paralegal approves every summary before a customer sees it; removing that step is a separate decision for later, based on the numbers
  • Excluded: drafting or redlining contracts, and answering legal questions

Terms

Invoices are due within 14 days. The code, prompts and evaluation set are yours. If phase 1 doesn't reach the target, you can stop there and pay only for phase 1.

Next step

Approve below and we'll send a short guide for choosing the 60 contracts.

A worked example for a fictional client. Every word and number is yours to change.

How to write each part

  1. Measure before you build

    Clients have seen AI demos that work once. Propose an evaluation set from their own data, with the answers their experts would give, and a score for every version. It turns “it seems good” into a number.

  2. Agree the target up front

    Write down what “good enough” means before any work starts (for example, every name, date and amount right in 95% of cases, and nothing invented). It decides whether the project continues after phase one.

  3. Keep a person in the loop

    For anything that reaches a customer, propose a review step first. Taking it away later is a decision for when the numbers support it, not something to promise on day one.

  4. Say exactly what happens to their data

    Which provider, whether it's used for training, how long it's kept, which region. This paragraph often decides whether the proposal gets past their legal team.

  5. Give the cost per use

    Clients worry about an AI bill that grows with usage. Measure the cost per document or request in phase one and put the figure in the report.

Questions people ask

What should an AI integration proposal include?

The task the AI will do and who checks it, how success is measured (an evaluation set and a target), the phases with a stop point after the first, data handling, running costs per use, the timeline and the price. The template above has each part written out.

How do you price an AI integration project?

Fix the price for the evaluation and prototype phase, so the client can stop there cheaply if the target isn't met, then quote the integration. The example above is about 90 hours across three phases before its optional extra.

How accurate does an AI feature need to be?

Accurate enough for the risk: a draft a person reviews can be useful at a lower score than something that goes straight to a customer. Agree the number with the client before building, and measure it on their own data.

Is client data used to train the AI model?

Not on most providers' business terms, but say which provider and terms you'll use, the retention period and the region, and confirm that the client's own customers have agreed to it.

Can I download this AI integration proposal template?

Yes: download it as Markdown for Google Docs, Notion or Word, or open it in Devposal and send it as a link your client approves online.

Related