SOW template · Data & AI
AI Chatbot Integration Statement of Work Template
An LLM assistant SOW that fixes knowledge sources, channels, intents, evaluation criteria and model costs, so an AI chatbot project does not chase an ever-moving accuracy target.
- 6
- Deliverables
- 6
- Exclusions
- 2 cycles
- Revision rule
Data & AI · Exclusion rule
Actions that change data in other systems (refunds, bookings, account changes).
Ambiguity fix
The chatbot will answer customer questions accurately.
The assistant will score at least 85% on the 200-item evaluation set agreed in Schedule D.
Best for
Engineering and data agencies adding a retrieval-augmented assistant to a website, product or support desk.
AI projects creep through expectations. A chatbot that answers 85% of questions well is followed by a request to make it 99%, then to add another knowledge base, then to let it take actions. This template defines knowledge sources, channels and intents, sets an evaluation method agreed in advance, and makes model and hosting costs the client's.
Copy any clause block below into your own SOW. Numbers, appendix names and tools are placeholders drawn from a typical engagement, so replace them with the counts from your proposal. Every clause is written to be read by a client, not used against one.
Deliverables
Count everything. A deliverable a client cannot count is a deliverable a client can expand.
- 2.1A retrieval-augmented assistant using the client's chosen LLM provider, indexing the 3 knowledge sources in Schedule A (help center, product docs, policy PDFs) up to 4,000 documents.
- 2.2A web chat widget for the client's website and a Zendesk handoff when the assistant cannot answer.
- 2.3Handling of up to 25 intents defined in Schedule B, including refusal and escalation behavior.
- 2.4An evaluation set of 200 question and answer pairs agreed with the client, with automated scoring.
- 2.5An admin view for reviewing conversations, flagging answers and re-indexing sources.
- 2.6Guardrails for personal data redaction in logs and topic restrictions listed in Schedule C.
Exclusions (out of scope)
Name what the engagement does not include. These are the lines your PMs will cite most often.
- 3.1Actions that change data in other systems (refunds, bookings, account changes).
- 3.2Voice channels, Slack, Teams and messaging apps.
- 3.3Model fine-tuning and training custom models.
- 3.4Knowledge sources not in Schedule A and writing new help content.
- 3.5LLM provider usage fees, vector database fees and hosting costs.
- 3.6Legal or regulatory review of assistant answers.
Revision limits
Prompt and retrieval tuning runs in two improvement cycles against the evaluation set. After the second cycle, further tuning is offered as a monthly optimization retainer.
Acceptance criteria
Acceptance needs a measurable test and a window. Without both, delivery never formally ends.
- 5.1The assistant scores at least 85% on the agreed evaluation set using the scoring rubric in Schedule D.
- 5.2Escalation to Zendesk works for every intent marked for handoff.
- 5.3Personal data patterns in Schedule C are redacted from stored logs.
Client assumptions & dependencies
Tie each dependency to a date so a late input has a known effect on the schedule.
- 6.1Knowledge sources are current and accurate. The assistant reflects the content it is given.
- 6.2The client provides API keys for the chosen LLM provider and Zendesk.
- 6.3The client's subject matter experts write or approve the evaluation answers.
Change control
New sources, channels, intents, actions, model providers or a higher accuracy target are changes. The agency issues a change order within two business days with hours, cost and any change in evaluation results expected.
These clauses are drafting guidance written from agency delivery practice. They are not legal advice. Agencies should have final MSAs and SOWs reviewed by counsel, especially for liability, IP, payment and termination terms, which this template does not cover.
Ambiguity check
The phrases that turn into free work, and what to write instead.
These are the vague lines we see most often in data & ai SOWs. Each one sounds reasonable at signature and becomes a dispute three months later.
Before
The chatbot will answer customer questions accurately.
Safer rewrite
The assistant will score at least 85% on the 200-item evaluation set agreed in Schedule D.
Why it matters: Accuracy needs a test set or it becomes whatever the last bad answer implies.
How to adapt it
7 steps from template to signed SOW.
A template saves drafting time. The protection comes from replacing every placeholder with the real counts, names and dates from your engagement.
- 1
Read the AI Chatbot Integration SOW end to end and mark every number, appendix and named tool that does not match your engagement.
- 2
Write the evaluation set with the client before the build. It is your acceptance test.
- 3
Pick the LLM provider in the SOW, or add a provider evaluation phase.
- 4
Exclude actions explicitly. Every client asks for them after launch.
- 5
Offer the optimization retainer up front so improvement has a home.
- 6
Run the ambiguity check: search your draft for words like all, any, ongoing, as needed, unlimited and timely, and replace each one with a count, a list or a date.
- 7
Have your counsel review the final MSA and SOW before signature, then load the signed SOW into TWINJAM so every request is checked against it.
Editable Word file
Get the AI Chatbot Integration SOW as a DOCX.
Every clause on this page in one formatted document, with numbered sections, appendix placeholders for your counts and a one-page change order form to attach to the SOW.
- 18 numbered clauses
- Revision and change control language
- Change order form with hours and rate fields
AI Chatbot Integration SOW.docx
Editable Word file with appendix placeholders
One email with the file, then The Profitable Agency newsletter. Unsubscribe in one click.
Related templates
- Data & AIData Warehouse Build SOWA data platform SOW that caps sources, models and dashboards, defines data quality responsibility, and stops a warehouse build from becoming an endless reporting queue.
- MobileReact Native App SOWA fixed-fee SOW for cross-platform iOS and Android builds that names every screen, caps device coverage, and keeps store submission and backend work from drifting into the base fee.
- E-commerceShopify Plus Migration SOWA replatforming SOW that fixes the data you migrate, the apps you replace and the redirects you map, so a Magento or WooCommerce move to Shopify Plus does not absorb every loose end in the old store.
Client asking for something outside the SOW already? Use a change order script to raise it warmly.
Signed the SOW? Let TWINJAM hold the line.
TWINJAM turns these clauses into boundary rules, flags Slack and Jira requests that cross them, and drafts the change order with the clause already cited.