The Founder as Proof
In January 2026 the business was making a loss, and I wrote every proposal, report and email by hand. Ten months later Claude runs across the whole business. Here are the numbers, including the ones that went the wrong way.
Christopher Duffy, Founder, IgniteAI Solutions · Updated 21 October 2026, figures to mid-October 2026
+299%
Invoicing, second quarter on first
+75%
October on September, a record month
24
New clients won, February to mid-October
~51 hrs
Saved a week, modelled since 31 August
~9×
Revenue this year vs all of 2025
28
Scheduled jobs running, 22 in the cloud
1,250+
Claude working sessions since March
15×
Google search impressions, January to September
2025, the first year of trading, closed at a loss. Revenue from January to 21 October 2026 is about 9 times the whole of 2025, read from Xero for the UK company, corrected for a bank refund posted twice to sales, before year-end adjustments and without the UAE company.
Since our June case study
In June I published a first case study, four months in. It made three claims: revenue had grown several times over against January, the system had produced 489 pieces of work in four months, and it was saving me five to nine hours a week. Four months later, here is how those claims held up.
What's new since June is the larger part of this study: a board, a full audit, a security incident and its fix, a second pivot, a second company in the UAE, a move across the world, and the move of the whole system into the cloud.
Background
IgniteAI trains small and medium businesses to use AI safely, and keeps them safe and progressing. I lead it after 23 years in UK Defence, with a small team of specialists in governance, training and development, and six university students on the associate bench.
Over ten months I rebuilt the way the business runs around Claude. Notion holds the state of the business: the pipeline, every engagement, every meeting and every task. A shared Google Drive holds every file. A Google Sheet holds the numbers, including the ledger this case study is built from. Claude sits across all three through connectors, runs on a schedule as well as when I ask it something, and follows about 55 skills I've written, each one a procedure for a job I used to do myself.
One rule sits above the rest. Claude never sends anything. It reads, drafts, files and updates records, and every email, proposal and invoice still goes out under my hand. Money is read-only and nothing gets deleted.
I'm publishing the numbers because most of what gets written about running a business on AI is either a demo or a promise. Some of these numbers are good, a few went the wrong way, and I've left those in, because a case study about measurement that hides its own bad months isn't worth reading.
What happened
The system was built in stages, and the order matters more than any single piece.
January and February: the starting point. The business was running at a loss. I'd started dictating instead of typing at the end of 2025, which turned out to be the habit everything else depends on, and in February Claude went into daily use for client work.
March to June: building, and the first lift. The first scheduled job ran on 3 March. Skills followed one by one, the knowledge base got its first entries in April, and in May the system was connected to the accounts so it could read invoices for itself. Client output went from almost nothing to dozens of documents a month, I won 14 new clients between March and May, and invoicing in the second quarter was four times the first.
July and August: two shocks at once. In July one large client took its work in-house. Over the same months I left the UK, moved house, relocated to the other side of the world and started establishing the business in a new country. Invoicing fell 78% in July. The system kept delivery running while I was carrying boxes, and by August invoicing had doubled again. Meanwhile the knowledge base was built out properly, a chief-of-staff agent started running the daily brief, and an advisory board of Claude seats was set up.
September: consolidation, a quiet failure and an incident. The two largest retainers of the year were won in September, and the business started billing retainers through its UAE company. On the system side, everything moved into one folder structure and half of the skill library, never used, was cut. The clean-up had a cost I didn't see at first: six scheduled jobs went dark for about a week, and nothing told me. In the same month the business had its first security incident, described below.
October: the audit, the pivot and a record month. On 1 October my chief-of-staff agent ran a full audit of the business, and I acted on it the same week. Most jobs moved into the cloud, an enquiry form went live, and a second agent, built on xAI's Grok, started publishing to X and Substack. Five new clients came in during the first half of the month, enquiries doubled, and October became the biggest month so far.
Growth
Growth is shown as an index rather than revenue. IgniteAI trades through two companies, one in the UK and one in the UAE, and the figures are net invoicing across both. March, the first month with a base, is set to 100, so 200 means twice March's invoicing.
Both companies, net invoicing. October is a record.
Source: the business ledger, net invoicing across the UK and UAE companies.
By quarter the shape is clear. The second quarter was four times the first, an increase of 299%. The third quarter fell 35% on the second, almost all of it in July, when a large client took its work in-house. Since July invoicing has risen three months running, and October, at more than twice March, beat June by 12% to become the record.
What the numbers show is that the system made the business resilient. It didn't stop a large client leaving, and it didn't stop a move across the world taking my attention for most of a summer, but invoicing dipped and recovered within a month, which it wouldn't have done when everything ran through me.
Clients
Clients invoiced, active engagements and new clients won. October to the 16th.
Source: the CRM and the business ledger.
Active engagements peaked at ten in April and are back at ten in October, and the two peaks are different shapes. In the spring most of the work was one-off: a training day, a discovery day, a short build. The system made it possible to say yes to all of them, and it was exhausting. Through the summer the work moved to retainers and phased delivery, each with weekly sessions, governance work and a stream of follow-ups. Since the October audit it's moving again, to fixed-price products, described below.
Demand
New enquiries, counted as organisations that contacted me first or were introduced to me, were in ones and twos a month through the spring, rose to two in July and three each in August and September, and reached six in the first half of October. Part of October's jump is the new enquiry form, and part is the content the system now produces every week. An enquirer in July told me Claude had recommended me, a channel I didn't build and can't yet measure.
Search is the slower engine underneath. The number of times the site appeared in Google results grew from about 1,100 in January to more than 16,000 in September, and clicks doubled over the same months. The average position improved from about 15 to under 10 in October. Most clicks still come from people searching for the business by name, and the pages bringing in new visitors are the practical guides: a comparison of Claude and Copilot, a guide to AI tools for small businesses, and the training page, all drafted through the system.
Productivity
I don't time myself, so the hours figure is modelled. Since 31 August every piece of work Claude does has been logged against a rate card I signed off: 300 minutes for a client document of five pages or more, 120 for a shorter one or an article draft, 30 for a meeting written up into actions or the morning brief, 25 for a substantial email and 10 for a short one. The rate card counts only Claude's part, and anything that doesn't fit is left out rather than guessed.
Against the rate card, with and without meeting write-ups.
Source: the work ledger, 317 logged pieces of work, 31 August to 18 October.
That's 317 logged pieces of work over seven weeks, an average of 51 hours a week, or 46 without meeting write-ups, which may overlap with the morning brief. Long client documents are 31% of the hours, articles 16%, building skills and automations 12% and meeting write-ups 10%. Email, where most people start with AI, is only 7%. The weekly figure swings from 30 hours to 66, and the low weeks are the ones where the work moved into rooms and meetings.
Output
Dated client documents went from none in January to 19 in April and 56 in June, and have run between 20 and 30 a month since. The early months are undercounted, because finished clients' folders were archived. Weekly, the count fell from 43 at the end of August to 11 in mid-October, partly because the first weeks were overcounted, and partly because the work changed shape: lots of short documents in September, a handful of long delivery documents by October. Meanwhile meetings with other people went from 9 a week to 24, emails I sent from 48 to 140, and tasks closed from almost nothing to 30 to 50 a week.
Claude usage
March from the 22nd. September and October count scheduled runs only, estimated.
Source: session logs to the end of August, scheduled-run counts after that.
Of the 893 sessions to the end of August, 530 were scheduled jobs and 363 were me asking for something. September and October add about 370 more scheduled runs, not counting anything I started myself. My ordinary Claude conversations over the past year come to 677: client delivery 30%, content 22%, building and fixing the system 18%, and strategy 12%. In one fortnight in September, 43% of my time in front of Claude went on fixing the system rather than using it. That's the hidden cost of building your own.
The knowledge base
The knowledge base is what stops Claude starting from nothing every morning. Notion holds the record, now 1,597 core entries linked to 102 companies, 326 people and 112 concepts, plus a research library of Claude practice added in October. A searchable copy is rebuilt from it, and above both sits a small set of hand-written pages for what must be right every time: what we sell, what it costs and who it's for.
Across all the stores, entries added each month were 152 in April, 57 in May, 188 in June, 710 in July, 1,015 in August, 146 in September and 826 in the first half of October. July and August were inflated by one-off backfills of past meetings and files, and normal intake was nearer 50 a month. September's low is the stall described above. October's rebound is mostly the new research library, about 800 entries, with meetings now going to their own record, 30 of them so far this month.
The rule that made the most difference is that automated jobs can propose changes to the hand-written layer but can't make them. A Thursday job compiles each client's week into proposed updates, marks anything two sources disagree on as contested, and I approve the changes on Friday.
The incident
In September the business had its first security incident, and it's the most important thing in this study for anyone thinking of running client work through AI.
A client needed a plugin installing. Instead of the clean built version, the folder I shared was the working folder the plugin is built from, 175 files of it. Inside it were the names of two other client firms, used as examples in skills and test sets, and a few first names. No contact details, no financial figures and no client documents or data went with it. The client spotted it himself, told me, and confirmed in writing that the files had been deleted. Access was revoked the same day.
The root cause was worse than the slip. Even the correct build would have carried names, because the master plugin used real clients as examples, and the build script removed names from a hand-kept list that was always one client behind. A second, smaller exposure to another client was found in the same review.
The fix is a check written in code. Nothing leaves the business until a script has read the exact file being sent, including folders, nested archives and office documents, and compared every name in it against a register of every client and their people. It fails closed: an unknown client or a missing register entry stops it, a planted test name has to be caught before any result counts, and there is no way to skip it. The master plugin now uses invented names only. Since 29 September it has run on every build, and the incident hasn't recurred.
An advisory board
A founder-led business has nobody to argue with. In July I gave it a board.
The board sits inside the system and every seat is Claude. There are thirteen: nine advisory seats, each written to challenge the business from a different discipline, a finance seat, a seat for an AI-led operator's view, and two staff seats filled by my own chief-of-staff and growth agents. The rule at the top of its brief is short. Staff seats report, advisory seats challenge, and I decide. None of the seats is a real adviser, and the brief says so on every page.
It was set up to meet every Monday evening on a schedule. It reads the pipeline, the books, the mail and the calendar, and puts each decision in front of me as a card with a recommendation and a go or no. Between late July and mid-August it produced 33 of those cards, from "raise this invoice this week" to "no client-facing draft survives 48 hours", which is now a standing rule. I overruled it once, when it called a situation red that wasn't, and I logged that too.
The board's biggest piece of work was a full audit of the business on 1 October, run by Woody, my chief-of-staff agent, on Claude Fable 5.1. It covered the prompts and skills the system runs on, the products, how far the business really runs on AI, and the people. It didn't spare me. A few of its lines:
"Work in this business moves when Chris touches it and waits when he doesn't."
It ended with ten decisions. The ones I acted on that week: a one-day Claude build for small businesses went live at a fixed price on 1 October, an older starter offer was retired on 2 October, a support plan replaced open-ended aftercare on 5 October, and a default of 14 days now applies to anything that would otherwise drift. On 16 October I stopped taking on new retainer clients, to put the time into packaged plugins and training for small businesses. Its recommendation to hand delivery to an associate is still open.
That's the second pivot of the year. The first, in the spring, was from one-off days to retainers, and the retainers carried the business through the summer and into a record October. The second is from retainers to products, because the audit showed that a business built on retainers still runs through one person. It's early, and the first two clients bought the one-day build within a fortnight of launch, and both are invoiced.
What changed in the system
Nearly every rule in the system exists because something went wrong, and the habit that paid off was asking each time whether the failure could be checked by code, because a written rule gets skipped on a busy day and a check that blocks the write doesn't.
Each morning, five of the 28 scheduled jobs run before I start: a pipeline sweep at 05:30, call preparation at 06:12, a money and pipeline picture at 07:30, the daily brief at 08:00, and a second pass at 09:20 that turns my answers into the operations brief. Weekly jobs cover the client sweep, prospect scans, overdue invoices, the knowledge base and this ledger, and monthly ones audit the system itself.
Where the business goes next
The system has turned into a business model, and it runs in the order clients actually buy.
Training brings them in. Accredited, hands-on Claude training on the team's own work, from half-day workshops to leadership days, with public masterclasses at the top of the funnel. Training is where most clients start, because it's the lowest-risk way to see what Claude does.
Uncertainty brings governance and discovery. Once people are using AI, the questions arrive: what's allowed, what's safe and what's worth doing. Governance setup answers the first two. A screener maps where each workflow's data travels, an AI policy is tied to that map, and an AI manifesto sets the approved tools, the use cases, the process for adding a new one and what happens after a breach. Insurers now read AI as part of cyber risk, so this is increasingly what a business needs to show its broker. The discovery audit answers the third question: interviews, process mapping and a review of existing systems, scored into a ranked list of use cases with the risk of each. One environmental consultancy's audit found 36 opportunities and a travel firm's found 16, from AED 24,000 (about USD 6,540) against AED 192,000 to AED 240,000 for a comparable review from a large consultancy.
The managed service recurs. The Claude managed service keeps clients working and safe: Claude issues resolved when they come up, protection against potential security threats as tools and connectors change, and ongoing support with the business's data.
Three routes take this to market at scale: Claude training for the UAE government in Abu Dhabi, a direct training partnership providing Firebrand's Claude training arm, and products built from the method, including the governance screener and agentic decision-making built on the client's own data.
The team behind it
Every member of the training team will be a Claude Certified Architect, Practitioner (CCA-P) and a UK AI Council Accredited Training Practitioner. IgniteAI is a Claude Certified Services Partner in Anthropic's Claude Partner Network and a UK AI Council Accredited Training Provider, and its partners include Anthropic, the UK AI Council, Supermetrics, Firebrand and the Dubai Business Women Council. I was named one of the Top 20 AI Leaders at CONFEX 2026.
What we recommend
For any small business thinking about doing the same, these are the things the numbers back.
The hours saved are modelled, not timed. Deliverables before June are undercounted. September and October Claude sessions are estimates of scheduled runs only. The website's analytics tool has been undercounting since late August, so the search figures come from Google's own search data. The Grok agent is days old. Growth is correlation with a clear mechanism, not a controlled test: Claude is one driver among several. No client has yet given written permission to be named, so no client is.
I'll publish an update at the end of the quarter with November's result.
Most clients start where this study ends: with training on their own work. Teams start with a Claude training day. A business of one or two people starts with the Founders plugin.