AI works best on first drafts and assembly: pulling approved information into a proposal template, summarizing the numbers you already track into a report, and sorting routine admin into the right queue. Give it only approved source material and a fixed template. Have a named person check facts, prices and commitments before anything goes out. Measure review time as well as drafting time.
The workflow at a glance
- The buyer question: An owner or manager spending evenings writing proposals, monthly reports and routine admin from scratch: what should we change first, and how will we know it worked?
- The current process: Each proposal starts from an old file, report numbers are copied by hand from several tools, and admin requests are handled one email at a time.
- What can improve: AI drafts from an approved source library and fixed templates; numbers come from the systems of record, not from the model; admin requests are triaged and routed with a draft response.
- Tools and access needed: Access to the proposal template and past approved proposals, the reporting data sources (read-only where possible), a shared inbox or request form, and an AI tool your business has approved under a written data rule. NIST’s generative AI profile lists confabulation (confidently stated false content) as a known risk to manage (NIST AI 600-1).
- Where people still approve decisions: A person approves scope, price and commitments in every proposal, interprets every report before it goes to a client, and signs off on any admin response that commits money or time.
- How to measure whether it works: Hours per proposal and per report before and after, the number of edits needed per draft, and factual errors caught at review.
Examples marked “Hypothetical example” are illustrations, not client results.
Generative AI is already a normal part of office work. Microsoft’s 2024 Work Trend Index, a survey of 31,000 people in 31 countries, found that 75% of knowledge workers use generative AI at work, and 78% of those users bring their own AI tools (Microsoft Work Trend Index, 2024). Your staff probably use AI already. What matters is whether they use it with the right inputs and checks. The steps below set those controls.
Step 1: Separate drafting from deciding
Sort the work into two piles. Drafting and assembly covers structure, boilerplate, summaries, formatting and pulling in known facts. Deciding covers price, scope, commitments, anything sensitive, and the final wording a client sees. AI helps with the first pile. The second stays with a person. NIST’s Generative AI Profile warns that these systems can produce confident but false output (“confabulation”), and it recommends defined human oversight (NIST AI 600-1).
Step 2: Build an approved source library
Most bad AI drafts come from bad inputs. Gather the material a draft is allowed to use: service descriptions, approved case studies, standard terms, team bios and past proposals you’re happy to reuse. Tell the tool to use only those sources, and to mark gaps instead of filling them. If a fact isn’t in the library, the draft should say “[confirm]” rather than guess.
Step 3: Take proposals from template to review
- Template. Fix the sections: the problem, the approach, deliverables, timeline, what’s needed from the client, and the investment.
- Intake. Capture the discovery-call notes in a consistent format. If you’ve connected inquiries to your CRM, pull them from the record.
- Draft. AI fills the template from the intake and the source library.
- Review. A person sets the price, confirms the scope, and checks every factual claim.
Hypothetical example (not a client case): After a discovery call, a consultancy’s notes are saved to the CRM. A drafting step fills the proposal template with the client’s stated goals and the matching service description, and leaves the price blank for the founder to fill in.
Step 4: Fix the report metrics and let AI write the narrative
Decide which numbers each report contains and where they come from. Assemble those figures automatically from your systems. Then let AI draft a plain-language summary of what changed. Two rules apply. Pull the numbers straight from the source system, and do not let the language model calculate them. A person adds the interpretation and recommendations. Ask the draft to quote the figures it describes so a reviewer can check them quickly.
Step 5: Triage and route repetitive admin
Inbox triage, meeting notes, status updates and research summaries are common candidates. A useful pattern is to classify each incoming item, draft a response or a task, and send it to the right person’s queue. Anything ambiguous goes to “needs a human.” Keep a log so you can see what was sorted and why.
Step 6: Set data and privacy rules
Decide which information can go into which tools. Because many employees bring their own AI tools (Microsoft, 2024), write down which tools are approved for client data. The FTC’s guidance to businesses is to know what personal information you hold, keep only what you need, and protect it (FTC). Check your AI vendor’s own terms on how it uses and retains data.
Step 7: Measure the time saved
Record how long each task takes today. After launch, measure drafting time plus review and correction time. A draft that takes five minutes to generate and forty minutes to fix hasn’t saved anything. Track how many drafts needed major changes, and use that to improve the template and the source library.
Where AI fits in each task
| Task | AI drafts or assembles | Person decides | Key control |
|---|---|---|---|
| Proposal | Structure, approach and service descriptions from the library | Price, scope, commitments | Gaps marked “[confirm]” |
| Client report | Narrative summary of figures from the source | Interpretation, recommendations | Numbers pulled straight from the system |
| Status update | Summary of task changes | What to send to the client | Approved before sending |
| Meeting notes | Summary and action items | Owners and deadlines | Attendee check |
| Inbox triage | Category and draft reply | Anything sensitive or unclear | “Needs a human” queue |
| Research | Summary with links | Conclusions | Sources cited and checked |
Checklist: AI for proposals, reports and admin
- Drafting tasks and decision tasks are listed separately
- An approved source library exists and is kept current
- Drafts mark missing facts instead of filling them in
- Prices, scope and commitments are set by a person
- Report figures come straight from source systems
- Approved tools for client data are written down
- Every client-facing output has a named reviewer
- Baseline task times are recorded, including review time
- Heavily edited drafts are tracked and used to improve the template
Related guides
Start with how to choose the first business workflow to automate. To roll these practices out across a team, read implementing AI across your team’s existing tools. For budgeting, see what business automation costs.
Research, reporting, document preparation and client updates are good first workflows for Grownup Business. Each project includes discovery, implementation, testing, a team handoff and agreed documentation, with high-stakes judgment kept behind explicit approval. Request a demo.
If Grownup also manages your Instagram, the same workflow can pick up inquiries that start there. Instagram Management is a separate, optional service. You do not need it to start any business workflow.