AVODA AI Consulting · Case study
Six days, USD 255, and a website that works
Before we sold forward-deployed AI engineering to anyone else, we ran it on ourselves. This is what it produced, what it cost, what it did not do, and what we would do differently.
Work delivered 14 to 21 July 2026. Measured to 21 August 2026. Every figure below can be checked against a live website.
The case study we could actually prove
Client work is confidential, which leaves every new consulting practice with the same problem: the proof is real and unpublishable. So we are publishing the one engagement where the client is us, the numbers are ours to disclose, and the site is public.
In July 2026 AVODA rebuilt avodagroup.org using exactly the method we sell: one person embedded in the operation, working with AI as an instrument rather than a novelty, building until the thing ran. It was not a demonstration. It was a real rebuild of a site the organisation depends on for applications, enquiries and credibility.
What we were starting from
The site was not neglected in the way people usually mean. It was maintained, it was online, and nobody inside the organisation thought it was a problem. That is the more common and more expensive condition, because nothing prompts anyone to look.
| What an audit found | Why it mattered |
|---|---|
| Homepage took 11.3 seconds on desktop, far longer on mobile | The audience is mobile-first and often on a metered connection. A page that slow is not slow. It is absent |
| The published programme fee was wrong by an order of magnitude | Applicants were reading a number nobody had charged in years and deciding they could not afford it |
| Zero programme pages carried real, current prices and dates | The most common question received by phone was answered nowhere on the site |
| Stale pages from 2022 and 2024 were still indexed | Search engines and AI assistants were quoting details that had not been true for two years |
| Zero internal links from articles to any programme page | The blog carried the traffic. It carried it nowhere |
| An active plugin was rescanning 185 posts on every page load | Nobody knew. It had been there for months |
Priced as external work the scope came to roughly 162 hours of specialist web engineering, technical search work, content editing and research: an estimated USD 17,150 to 43,350 and a delivery schedule of two to three months. That is an estimate of what a supplier would bill, not a quotation anyone issued, and it is labelled so nobody mistakes it for an invoice.
What forward-deployed actually looked like
Not a strategy phase followed by an implementation phase. One person, inside the systems, shipping every day, with AI doing the volume work and the human doing the judgement work. The distinction between those two is the whole method.
What the AI did
- Audited more than 200 pages and posts far faster than a person could open them
- Wrote and rewrote page copy against a defined brief, in draft, for a human to cut
- Generated the schema markup, redirect rules and performance configuration
- Found the plugin rescanning on every page load, unnoticed for months
- Compressed and converted 90 images and produced the rules to serve them
- Held the whole site in view at once, which a human working page by page cannot do
What the human did
- Decided what was true. Every price, date and claim was checked against a source, and several the AI proposed were wrong
- Refused the blanket find-and-replace that would have been faster, because it would have published false statements on live pages
- Chose what to delete. Twenty-three stale pages were retired, which is a judgement call with consequences
- Carried the relationships. Confirming a date with a faculty member is not a task you delegate to a tool
- Owned every published word
The work did not become easier. One person working full days produced the output of a small team, and the days were long.
Measured, with the caveats attached
| Measure | Before | After |
|---|---|---|
| Homepage load, returning visitor | 11.3 seconds | around 0.5 seconds |
| Programme pages with real prices and dates | 0 | 7 |
| Published programme fee | Wrong, by an order of magnitude | Correct, and public |
| Internal routes from articles to programmes | 0 | 63 |
| Email authentication | SPF only, unsigned | SPF and DKIM live |
| Database size | 630 MB | 105 MB |
| Machine and agent readiness checks passing | Not measured | 13 of 13 |
What changed in the traffic
Over the first thirty-day window that included the rebuild, against the preceding thirty days: sessions +83%, pageviews +133%, users +113%, and pages per session 1.28 to 1.63.
The honest caveat on those figures
Only the final week of that thirty-day window was after the rebuild. Attribution is therefore not clean, and we say so rather than presenting the whole increase as caused by the work. The figure we would defend is the pages-per-session movement, because that is a behaviour change consistent with pages that now load and answer the question. Anyone presenting a chart like this without that paragraph is selling you something.
What AI did not do
Five things, each of which cost real time and none of which the tooling solved. If you are budgeting for an engagement like this, budget for these.
1. It did not save the person’s time
Six working days were worked, and they were long. What changed is what a working day produces, not how hard the day is. Any proposal that promises an easier week is describing something other than this method.
2. It got facts wrong, confidently
Prices, dates and claims proposed by the AI had to be checked against source documents, and several were wrong. The most important refusal of the whole engagement was declining a blanket find-and-replace that would have been much faster and would have published false statements about a programme on live pages. A tool cannot make that call.
3. It could not decide what was true
The organisation’s own documents disagreed with each other on several figures. Reconciling them was human work involving source audits and conversations, and it took longer than the rebuild’s technical work. No amount of tooling resolves a contradiction between two internal documents.
4. It could not carry the relationships
Confirming a date with a faculty member, agreeing what a programme would be called, getting a decision on what to retire. All of it human, all of it on human timescales, none of it compressible by a tool.
5. It did not remove the need to know the work
The reason six days was enough is that the person running it already understood websites, search, the organisation and the market. Given to someone without that, the same tooling produces a great deal of confident output and no way to tell which parts are wrong. This is the finding we most want a prospective client to take away, because it decides whether an AI engagement works in their organisation or produces expensive noise.
AI multiplied the output of someone who knew what they were doing. It did not substitute for knowing.
What transfers to your organisation, and what does not
Transfers
- One embedded person beats a phased vendor engagement on both time and cost, until the work needs a genuine team. We will say so when it does
- The bottleneck is deciding what is true, not building
- The largest wins were unglamorous: correct prices, pages that load, links that go somewhere
- One component is worth more every year. Find it early and build around it
Does not transfer
- The six days. This organisation is small, the decision-maker was in the room, and there was no procurement cycle
- The single operator. It worked because one person held the whole site in their head. Beyond a certain size the method changes
- The absence of approvals. Nobody had to sign off copy. Where they do, plan for it rather than discovering it in week two
What we would do differently
- Fix the measurement before the site. The traffic caveat above exists because the analytics window straddled the work. A clean baseline measured a fortnight earlier would have cost almost nothing and made the result defensible
- Name the maintainer on day one. Corrected prices are the shortest-lived thing built, and ownership was agreed after the fact rather than before
- Reconcile the organisation’s own figures first. That work turned out to be larger than the rebuild and would have been better done before rather than during
How to start
A scoping call is thirty minutes and costs nothing. We look at the operation and tell you honestly whether AI helps, and where. It regularly ends with us saying not yet, which is cheaper for both sides than the alternative.
Free tools, no email needed to see your result: an AI readiness assessment, a value calculator and a vendor scorecard. Templates, the cost baseline and a board briefing deck are on the AI resources page.