An AI agent for small business that drafts the work and waits for your yes
Discovery comes first. Then you get an agent tested on your own past requests, with approval checkpoints, a full log, and an off switch.
What would an agent do in a real business?
The same kind of request lands in your inbox in a different shape every time. Someone reads each one and keys it in by hand. Picture a 50-person freight brokerage. Quote requests pour into a shared inbox all day. Origin, destination, equipment type, pickup date, and weight show up anywhere. In the email body, in an attached spreadsheet, or three replies deep in a forwarded thread.
A broker reads each one and keys the details into the TMS by hand. Then looks up recent rates on the lane and types a reply. On a busy afternoon the requests stack up. A shipper books with whoever answered first.
An agent here would read each request and pull out the load details. It would find comparable rates in the brokerage's own history and draft a quote. The broker checks the number, adjusts the margin, and sends it. The agent never prices, sends, or books on its own.
Why does discovery come first?
An agent is only as reliable as what it reads and the rules it follows. Before pricing a build, Raquel needs to see four things. How do requests arrive? Is past rate data complete enough to rely on? Which systems would the agent reach? What would your brokers call a good draft? None of that shows up on a 20-minute call. The floor price starts the conversation. Your fixed build quote comes out of discovery.
How does an agent project run?
- 01Discovery
- 02Data and access review
- 03Scope and acceptance criteria
- 04Pilot build and evaluation
- 05Launch approval and handover
All ten steps, in detail
- Discovery sessions with the owner and two people who handle the requests today.
- A sample of recent requests gathered, including the ugly ones.
- The data the agent would rely on, such as rate history and customer records, reviewed for gaps and duplicates.
- System access mapped: inbox, TMS, rate sources, and whether each needs read or write permission.
- A fixed build quote, with acceptance criteria in your terms. For example, the share of test requests where the draft gets lane, equipment, and date right.
- A spec stating what the agent reads, what it drafts, where it stops, and who approves.
- The agent built by a contracted technical delivery partner, against the spec Raquel writes and reviews.
- An evaluation on a test set of past requests, scored against the acceptance criteria.
- A shadow run, where the agent drafts alongside your team on live requests while nobody uses its drafts.
- Launch approval from your decision owner, then handover documents, a team walkthrough, and a 30-day review.
What do you walk away with?
- Discovery findings with a go or no-go recommendation
- Data and access review
- Fixed build quote and written acceptance criteria
- Working agent with an approval step before anything goes out
- Evaluation results on the test set
- Shadow run notes
- Logging and rollback plan
- Handover documentation and team walkthrough
- 30-day review
What has to be true before anything gets built?
- Approved data sources: a written list of exactly what the agent may read, and nothing beyond it
- A named owner at your firm who answers questions and signs off
- An access and security review, granting the agent only the permissions it needs
- Acceptance criteria agreed in writing before the build starts
- Logging of every request, every draft, and every approval
- A rollback path, so your team can switch the agent off and return to the manual process the same day
What you get, what this is not, and what you need to bring
What you get
- An honest answer, before you pay for a build, on whether an agent is the right tool at all
- An agent tested on your own past requests before it touches a live one
- A record of every draft and who approved it
What this is not
- No agent that sends, prices, books, or commits on its own
- No guaranteed ROI, revenue, or win rate
- Not a company-wide AI rollout
- Not included in the founding cohort discount
What you need to bring
- A decision owner with authority to approve launch
- A sample of recent requests, with outcomes where you have them
- Access to the systems the agent would read
- One or two experienced staff to review drafts during the shadow run
Five questions that set the price (answered in discovery)
- How many request formats must the agent read: email body, attachments, forwarded threads, web forms?
- How many systems does it read from, and does it need to write to any of them?
- How complete and consistent is the data it would rely on?
- How strict are the acceptance criteria, and how large is the test set?
- How many people and approval steps sit in the loop?
Price
- What moves it
- Request formats, connected systems, data condition, evaluation depth, approval steps
- Payment
- Paid discovery ($1,500) first; the build is then quoted at a fixed price and paid 50% at kickoff, 50% at delivery
What if it isn't worth it?
Discovery ends with a go or no-go recommendation, before you pay for a build. If the answer is no, you keep the findings.
Later on: monitoring and optimization is available after launch for $400 to $1,200 a month.
Questions people ask before booking
Why can't you quote the agent today?
The price depends on things nobody can see yet. How many shapes do the requests come in? How clean is the rate history? How many systems must the agent reach? A quote without those answers would be a guess padded for risk. Discovery replaces the guess with a fixed number.
What if discovery shows an agent is the wrong answer?
Then that's what you hear, along with what would work better. Sometimes it's a smaller fixed-range build like a workflow sprint. Sometimes the data needs cleaning first. Either way you keep the findings.
Will the agent ever send a quote without a broker?
Not in anything Raquel scopes. A quote commits your business to a price, so a person approves every one before it goes out. The log shows who approved what and when.
Who owns the agent once it's live?
You do, once the final invoice is paid. It runs in accounts you control, and the documentation is yours. Any AI model or software costs are billed to you directly. Switch it off, and your team falls back to the manual process without losing anything.