AI automation for your business, built on your data
We build one working AI tool that takes a repetitive job off your plate: invoice emails into the books, incoming requests sorted with a reply drafted for approval, the weekly report that builds itself. It runs on your own data, in the tools you already own, is tested on your real past cases, tuned for 30 days, and handed over with documentation. Scope and quote in writing, before anything is signed.
Blue Peak is a data firm. AI is the newest thing a business does with its data, and an automation is only as good as the records, documents and rules it runs on. That is where most AI efforts stall, and it is where every build of ours starts.
One job, automated properly, not a pilot that never ships
We named the firm for data because everything a business does with its information is data work, and AI is the newest part of it. An automation that reads a messy export gives a confident wrong answer. One that reads a clean source, under a rule that says what may go in, with a person approving what comes out, gives a useful one. The difference is rarely the AI. It is the data underneath and the rules around it.
So a build with us starts by mapping the data the job touches: which mailbox, which spreadsheet, which fields, what each one means, what must never go into an AI tool. Then we build the tool in software you already pay for (Power Automate, Copilot, Claude or ChatGPT, whichever fits), test it against 20 of your real past cases with the accuracy written down, add a human approval step and a log, run it beside you for 30 days, and hand it over. Nothing we build sends, pays, deletes or decides on its own.
We do not resell licences, administer your Microsoft tenant or replace your IT provider. We are the builder, and we are glad to work beside the IT firm you already use.
Small businesses say the problem is trust, not access. In the January 2026 wave of Intuit QuickBooks' 2026 AI Impact Report (2,222 US businesses surveyed), the top barriers named were privacy and security (36%), not knowing what AI can do (28%) and concern about bias or errors (26%). Source: Intuit QuickBooks, 2026 AI Impact Report.
Businesses
Owners and office managers of firms with roughly 10 to 250 people: accounting and insurance offices, engineering and architecture firms, property management, staffing, trades and distribution back offices, and nonprofits. Anyone with one repetitive, document-heavy job that eats a person's week.
Public agencies
Cities, counties, school districts, police departments and special districts with a monthly packet somebody builds by hand, a records request queue, or a shared inbox that needs sorting. Blue Peak is a SAM.gov active small business, UEI PRPBEM9HEND3, CAGE 15CS5.
The repetitive work a small business recognizes
Each of these is one build. Every one keeps a person in the loop before anything leaves the firm, and every one starts from data you already have.
- 01Invoice emails into the bookkeeping sheet, with a one-click approval before anything is recorded.
- 02Incoming requests sorted by type, each with a draft reply waiting for a person to approve.
- 03Long documents summarised the same way every time, contracts, proposals and RFPs, with page citations you can check.
- 04The weekly report that builds itself from the export you already run, in the layout you already use.
- 05Intake forms and web enquiries into the CRM, cleaned, deduplicated and tagged.
- 06Quotes drafted from your price sheet and template, checked by the person who signs them.
- 07Meeting notes into tasks with owners and dates, open questions listed instead of guessed.
- 08Staff questions answered from your own handbook and procedures, with the source paragraph quoted.
Worked example: invoice emails into the bookkeeping spreadsheet, with approval
Example design, not client work- 01
An invoice arrives by email
To the shared accounts mailbox, as a PDF attachment, like today.
- 02
A flow picks it up
Power Automate watches the mailbox and runs on each new attachment. Nothing else changes for the sender.
- 03
AI reads the PDF
Vendor, invoice number, amount, due date and PO reference are extracted. Anything it cannot read with confidence is marked "check".
- 04
A row is added, status Pending
To the AP spreadsheet or SharePoint list you already keep, with a link back to the original email and file.
- 05
The bookkeeper approves or sends it back
An approval card in Teams or Outlook shows the extracted fields beside the PDF. One click approves; one click returns it with a note.
- 06
Filed and logged
Approved rows are marked and the file is moved to the right folder. Every step, every value and every decision is written to a log you can audit.
Built in your Microsoft 365 tenant with the licences you already have where possible. Tested against 20 real past invoices before it goes live, with the accuracy written down. Nothing is paid, posted or deleted by the automation; the person in step 05 is the only one who decides.
One Workflow, Automated. Scope written down before anything is signed.
Two sizes of the same build. Each is its own written scope and its own acceptance, and we quote it in writing once we know the job.
One Workflow, Automated
One repetitive, document-heavy job from the list above, built end to end in the tools you own: a Power Automate flow, a Copilot Studio agent, or a Claude or ChatGPT project, with a human approval step before anything leaves the firm.
The data step. The build starts with a map of every record, file and mailbox the workflow touches, and what each field means. Nothing is automated until that map is agreed and the source is clean enough to trust.
- A one-page process map of the job as it runs today
- The working automation, with the approval step, guardrails and a log of every action
- A test against 20 of your real past cases, with a written accuracy check
- A run guide and a recorded handoff session
- 30 days of tuning after go-live
About four to six weeks of build time from a scheduled start. We take one build at a time, and the start date is written into the scope. Files, flows and documentation are yours.
Single Connection in Microsoft 365
The narrow version: one Power Automate flow or Copilot Studio agent that connects two systems you already have in Microsoft 365 (a mailbox and a list, a form and a spreadsheet, a folder and a Teams channel), with an approval step.
The data step. The same map, for two systems instead of a whole job: which fields move, what each means, and what stays out.
- The process map for the two systems
- The working flow or agent, with the approval step and a log
- A test against 10 of your real past cases, written up
- A run guide and a recorded handoff session
- 14 days of tuning after go-live
About two to three weeks of build time from a scheduled start. The usual first build: it proves the pattern on your data before you commit to a larger one.
After either build, AI Desk: a monthly plan with five hours of tuning, new prompts and questions by email, month to month, cancel any time. Offered when there is room on the calendar.
No licence resale, no hosting fee, no seat count from us. Every engagement starts with a short written intake rather than a sales call, and ends with files you own. Each is its own invoice, Net 30, on acceptance.
Two smaller pieces, for the rules and the people
An automation lives inside a business that has rules for AI and people who can check its output. If you do not have those yet, these two are how we add them. Both have a free sample you can read first, written by Blue Peak and labelled as a sample, not a client document.
AI Use Policy and Readiness Check
A written AI use policy in plain English (approved tools, what data may and may not go into them, human review, disclosure, incidents, training), a one-page staff version, a short anonymous staff survey, a readiness scorecard of your tools, licences and where the sensitive data sits, the five best first uses of AI for your firm, and a recorded 20-minute leadership walkthrough. Mostly in writing and recorded video.
A four-page sample AI use policy for a small business, drawing on NIST guidance (the AI Risk Management Framework, NIST AI 100-1, and its Generative AI Profile, NIST AI 600-1) and the FTC's guidance on AI claims, all cited in it. The document says on its face that it is a template and not legal advice.
Team AI Workshop and Prompt Playbook
A written intake that collects three to five real tasks per role, one live two-hour hands-on session (online or on site) where every exercise uses your tasks and your approved tool, a role-based playbook of about 25 tested prompts with your data rules printed beside them, the recording, two weeks of questions by email and a 30-day check memo. Every exercise starts with what to keep out of the prompt and ends with checking the output against the source. A repeat session for a second group can be added.
A two-page sample written by Blue Peak: five prompts for everyday small-business tasks, the rule each one teaches, and the data rule. It shows the kind of prompts a client's workshop playbook contains. Copilot, ChatGPT, Claude and Gemini all take them as written.
Our own work, in the open
We have no client case studies to show you yet, and we will not invent any. What we can show is what we have built for ourselves with the same tools and the same review discipline we bring to a build. Each of these is live.
Blue Peak Bid Board
A subscription service we built and run. AI reads government solicitations that run to hundreds of pages and quotes the sentence that decides eligibility, with its page number, so a person can verify the call in seconds instead of reading the whole document. The sample board is public.
Open the sample boardA multi-agent proposal system with hard gates
The system behind our own government bids: specialist agents triage solicitations, research buyers, draft and red-team proposals, and scripts block anything that carries a banned claim, a wrong sum or AI-tell phrasing before a person reads it. It is how we know which review steps a working AI system needs, because we learned them on our own work.
How we workPolice crime statistics, demo to Power BI page
A working crime statistics demo computed from public Olathe Police Department records, and the Power BI page built from it with Microsoft's own AI authoring tools, validated and screenshot-checked before a human read it. Olathe is a public data source for the demo, not a client.
See the page and the demoA narrated data video, built from code
A narrated walkthrough of our crime statistics demo, built from public data by a pipeline, with short demos scheduled weekly from October: a written script, a synthetic voice, screen captures and captions assembled without a studio. The same pipeline produces a recorded how-to for each step of a client's automation in minutes.
Watch on YouTubeFour phases, the same as every Blue Peak project
The same four phases as our analytics work (how we work). From a scheduled start, one build at a time, a Single Connection runs about two to three weeks of build time and a full One Workflow build about four to six. The start date is written into the scope.
Discovery: map the data and the steps
A short written intake, not a sales call, plus a handful of real past cases. We map the job as it runs today, every record, file and mailbox it touches, what each field means, and what must never go into an AI tool. You get a written scope and a written quote before anything is signed. If a feature you already own would do the job, we say so.
Solution design: tool, approval step, test set
The tool choice (Power Automate, Copilot Studio, Claude or ChatGPT, and why), where the human approval sits, what gets logged, and the 20 past cases the build will be tested against, agreed in writing before anything is built.
Implementation: build and test on 20 real cases
Built in your tenant or workspace, against your real data, reviewed with you as it goes. Then run against the 20 past cases with the accuracy written down, so you know the error rate rather than hoping. One revision round is in the price; more are fine and are scoped as they come.
Optimization and handoff: 30 days of tuning, then it is yours
It goes live beside the old way for 30 days while we tune it. Then a recorded walkthrough, the run guide and the documentation are handed over, with no licence back to us. If you never call us again, it keeps working; if you want a hand each month, AI Desk is there.
Questions owners and managers ask
Which tools do you use, and when does each fit?
The ones you already pay for. Power Automate when the job moves data between Microsoft 365 systems (mail, SharePoint, Excel, Teams, Forms) on a trigger. Copilot Studio when staff need to ask questions of your own documents inside Teams. Claude or ChatGPT (business or team plans) when the job is reading, summarising or drafting from long documents, often called from a flow. Many builds use two of them. We do not resell licences and have no reason to steer you to one vendor.
What happens when it makes a mistake?
It will, sometimes, which is why nothing we build acts on its own. A person approves before anything goes out, anything the AI could not read cleanly is flagged rather than guessed, and every step, value and decision is written to a log. The 20-case test before go-live tells you the error rate in writing, and the 30 days of tuning are for the cases it gets wrong.
Who owns what you build?
You do. The flows, agents, prompts, test set, run guide and documentation are yours at handoff, inside your own accounts, with no licence back to us and no dependency on us to keep running.
What does it cost to run each month?
Your own licences, and nothing to us unless you choose AI Desk. Many flows run on what a Microsoft 365 business subscription already includes; some connectors and AI features need a premium Power Automate or AI Builder add-on, which is a published per-user Microsoft price we name in the scope before you commit. General assistants run about $20 to $25 a seat a month on business plans (ChatGPT Business, Claude Team), and Microsoft 365 Copilot Business is $21 a user a month at list for firms under 300 seats, and Microsoft runs promotions (Microsoft). Vendor list prices as published; the scope states the exact monthly figure for your build.
Where does our data go?
Into your own business accounts and nowhere else. We work inside your Microsoft 365, Google Workspace, ChatGPT Business or Claude Team accounts, whose business terms do not train on your data by default, and never through personal tools of ours. The data map names what never goes into an AI tool, and the engagement letter says the same.
Do we need Copilot licences first?
No. Most first builds run on Power Automate and a business-plan assistant you may already have. Where a Copilot or premium licence would make a difference, we say which one, what it costs per seat and what it would change, and you decide.
Can a public agency buy this?
Yes. Both builds sit under most small-purchase thresholds, each is a separate written scope and invoice, and we are a SAM.gov active small business (UEI PRPBEM9HEND3, CAGE 15CS5). If your purchasing office needs a quote with a statement of work attached, say so and it is yours.
Who does the work?
Blue Peak's senior team, and the consultant who scopes your build is the one who delivers it. Our lead consultant brings twelve years in production analytics, most of it in the Microsoft stack, and uses AI daily in the work you see on this page: multi-agent pipelines with review gates, Power BI pages authored with AI tooling, reusable prompt workflows that generated database rules at a national telecom, and analysis of support cases to improve an AI chat service at a consumer electronics business. That record was earned before Blue Peak, so we describe it as our people's experience rather than the firm's past performance, on purpose.
Tell us the one job you would hand to AI first
Name it in a line: the inbox nobody keeps up with, the invoices typed in by hand, the report that gets rebuilt every Monday. Add a rough figure for how long it takes today.
You get an honest read back within one business day: whether it is a Single Connection or a full build, what it would cost, and whether it is worth doing at all.