Common questions

Ordinary automation moves structured data between two systems along one fixed path. It stops when the input changes shape, such as a supplier who moves the total to a new spot on the invoice. An AI agent reads input that arrives unstructured, such as a PDF invoice, an email or a scanned form, picks the rule that applies, acts in your system, and passes the unclear cases to a named person. We build both. The Blueprint says which one each row needs. A row that a fixed rule can handle gets plain automation, because that costs less to build and run.

An agent will get some cases wrong. Each agent has a confidence threshold you set, and a case below it goes to a named person with the reason attached. Before go-live the agent runs in parallel with your team. We measure its error rate against theirs and print it on the Blueprint row. The log shows what the agent read and what it wrote, and the agent's owner can pause it. We measure again at 30 and 90 days. If a result lands below the forecast, we tell you the week we find it.

You decide that for each action, and the Blueprint row records it. On invoice matching the agent posts a clean match to the approval queue. Your controller still approves the batch. Payments, customer offers above your limit, and changes to pay records stay with a named approver. We do not build agents that approve spend, change bank details, release a payment, or sign a contract.

The agents work in the systems you already run: the ERP or CRM, the payroll platform, the shared drives, email, and Slack or Teams. Each agent gets its own service account with the narrowest permissions its task needs. Your IT team can see that account and revoke it the same day.

We start from the workflow and its numbers: how many cases a month, how many minutes each, and how often your team has to redo one. Each Blueprint row forecasts what an agent returns against those numbers, and the agent runs in parallel with your team before it takes live work. We measure it against the forecast at 30 and 90 days. A row whose return does not cover its build cost says so in the Blueprint, and we advise against building it.

The Audit lists the tools you already license, and the Blueprint says which rows one of them can carry. If a feature in your ERP or your current vendor can handle a row, the row says so, and you can give that work to them. The forecast on each row stays yours, so you can hold whoever builds it to the same number.

We quote the Audit as a fixed fee after the discovery call. The fee follows your headcount and the number of departments in scope, so an audit of four departments at 300 people costs less than one of nine departments at 2000. The Blueprint then shows the payback on every item before you spend anything on a build.

The Audit runs two to four weeks, depending on how many departments you put in scope and how fast we get calendar time with your people. The Blueprint lands one to two weeks after that. We scope build items one at a time, and we expect most to take two to eight weeks.

We need read access to the systems that hold the work: the ERP or CRM, the payroll platform, and the shared drives where the spreadsheets live. Add about 18 hours of staff time across a two-department audit, most of it 45-minute interviews, plus a sample document for each process we map. We sign your NDA, and we work under read-only credentials until a build needs to write. Each agent then gets its own account, with the narrowest permissions its task needs.

We report the headcount effect the numbers show, up or down, on every row of the Blueprint. An agent takes the routine cases in a workflow, and the people who run it today keep the exceptions, the approvals and the vendor calls. The Audit also lists the work that has sat in a queue for months because your team had no hours for it, so you can see where the returned hours could go before you change a role. We will not write a document to justify a decision you have already made.

You have paid for the Audit and the Blueprint, and you owe us nothing further. You keep the process inventory and the forecast with a build cost on each row, and you can hand any of it to your own team or another contractor. We would rather give you a page you decline than sell a build the payback does not support.

We build the AI agents and the automation ourselves. The same people who ran the Audit specify each build, and we bring in engineers when a job needs a skill we do not hold. You sign one contract, and one team answers for the forecast we wrote.

Yes, through our small-business team. The full Audit and Blueprint are built for companies of 200 people and up, where the volume carries the forecast. Under about 50 people a separate team runs a shorter engagement sized to your headcount, with the same agents and the same controls. Book the same discovery call and say your size, and that team takes the call. The starter path sets out the steps, from the mapping session to your first agent.

You tell us, and within 24 hours you have the files and complete control: the agent code and prompts, the runbooks, the logs, and the accounts each agent signs in with, transferred to a person you name. The agents keep running while the handoff happens. Nothing we build depends on a platform only we can operate.

In the Audit, IT shares about ten hours with your sponsor, finance and a coordinator, most of it on granting read access or running exports. If read access has to clear a security review first, we start from saved reports your team already runs. In the Build, IT creates one service account per agent with the narrowest permissions its task needs. Agents go live one at a time, so IT handles one account request per go-live.

We do not build without a forecast, because the forecast is the number we measure the build against at 30 and 90 days. If you know the workflow, put its department alone in scope. The fee follows the departments in scope, so a one-department Audit costs less and runs at the short end of two to four weeks. The Blueprint then prices your workflow beside the others in that department, and you still pick what gets built.

In the Audit we record a baseline for each workflow: cases a month, minutes per case, the error rate, and days from trigger to done. At 30 and 90 days after go-live we measure the same figures from the agent's log and your system's records, and set them beside the Blueprint row. You see forecast and actual side by side for each agent, including any that land short.

Not when your tools run in the cloud. Your agents run on an always-on cloud server and sign in to tools such as Gmail or QuickBooks Online through each tool's own permission screen, which you approve and can revoke. If a workflow needs files kept on one of your machines, we install a small connector there with you present, under its own login. We take remote control only to install or fix that connector or a desktop-only program, in a session you book and watch, and no remote-access tool stays behind.

Your agents run on an always-on cloud server, so your laptop can be off while they keep working in your cloud tools. For a solo business or a company under about 50 people we host that server, and the monthly retainer covers it. A larger company can run it in its own cloud account instead. The one exception is a connector on an office machine, which needs that machine on when the agent reads its files.

On that session we see only what you share on your screen, the same as any screen-share call. We never ask for your password: each tool's own sign-in screen handles that, and it goes to that company, not to us. No standing or unattended access, ever. The session runs only when you book it and watch it, and it ends the moment you close it. On the implementation package your data stays in your own Claude or ChatGPT account the whole time.

You pick where it runs: Claude or ChatGPT, our hosting, or your own hardware or cloud. You also pick how it starts. A standard rollout sets up the workflow and goes live on your say-so. A controlled rollout runs it in a test environment or on one dedicated machine first, on one team or one workflow at a time, read-only before it can write, with a named owner who can pause it before it ever touches live work.

The context session runs 45 to 60 minutes, after the mapping session and before go-live. You walk through three to five real past threads, with anything personal stripped out first. You answer how you open and close a message, your negotiation positions and walk-away points, and who gets a fast reply versus an escalation. We write up a style and negotiation profile your agent works from. It never holds a password, a personal message unrelated to the business, or another person's private data.