Of the 32 AI programmes in August's employer-adoption sample, 31 involve work inside the business. None point at customers alone.

1. Three editions point to the same place: work inside the organisation

The August employer-adoption sample contains 32 AI programmes. Twenty-three pointed the AI at internal work only, eight at both internal and customer-facing work, and one was too vague to call. None were customer-facing only.

Across the three editions:

That is 89 of 92 classifiable programmes across three months. This is now a recurring pattern in the ads we captured.

June's figure is restated. Every June and July ad has been re-read and classified against the same rules used for August, and the classification is now recorded on each record rather than held in a summary. June was previously published as 29 of 29; the re-read found four more June ads describing a programme, three of which were too vague to classify by audience. All 30 classifiable programmes involved internal work, so the finding is unchanged.

One caveat belongs right here rather than in a footnote. We search a different mix of terms each month, and that shapes what comes back. The adoption and enablement query family produced 27 of the 46 August ads, and an enablement role is internal almost by definition. July found the same thing on a mix that included operations- and technology-manager titles.

Internal doesn't mean a back-office-only project. August's mixed roles covered customer follow-up, recommendations, support, community services and digital experience as well as finance, reporting, operations and staff productivity. The common element was internal ownership: someone inside the organisation was expected to choose the work, implement it, help people use it and carry the result.

One obvious question: is this an artefact of who we left out? We set aside 13 August ads under the Index's supply-side rule: 12 from consultancies or AI product businesses, and one role that combined internal and client delivery. Those 13 contain no customer-facing-only programme either. Across all three editions the set-aside group does lean further toward the customer than the included one — four of its 30 programmes point the AI at customers alone, against three of 97 on the adopter side — so where customer-facing AI shows up in these samples, it is more often being built by the businesses selling AI than by the businesses buying it. In August the line makes no difference at all: the count is zero on both sides of it.

2. There is no single AI stage

For the third edition, the ads describe work at all four stages.

June's and July's rows are restated on the re-read described above; they were previously published as 9/8/4/8 and 10/7/7/7. Do not read the columns down the page. The three rows don't count quite the same population, the search terms differ every month, and where a single ad sits between two stages is a judgement call — enough of them moved on re-reading to swamp any real month-to-month movement. What the table shows is that all four stages appear in all three months.

The August briefs ranged from reviewing systems and setting a roadmap, through proofs of concept and minimum viable solutions, to rollout, early-life support and production monitoring. A single-office law firm was starting with a workflow review in the same month a construction group described itself as “mid-build on a full AI and automation layer” and a national services group already had Python agents deployed as containerised services, and wanted them instrumented for cost, latency and quality.

Three months in, all four stages are still turning up at once. No stage in this sample is one the others have left behind.

3. The job doesn't end at go-live

Several August ads hand one person the whole loop, and it has the same shape each time:

Choose a problem. Establish the baseline. Build or buy the change. Help people use it. Check the number again.

Not every ad names all five steps. The ones that put a number on the work name most of them, and they make a single person accountable from the first decision through to the measurement afterwards.

A technology leadership group advertised a six-month contract to measure how much of its leaders' time was going on operational overhead, and cut it from 20% to 10% inside six months. A construction business asked for $500,000 of documented value in year one and 70% active users measured 90 days after each solution went live. A consumer-products business of about 180 people wanted someone who would pull the licence report, remove the seats nobody was using, and rewrite the internal guidance when a model update broke a worked example.

Others are less numeric but the same shape: post-implementation reviews checking benefit claims against the original business case, and monitoring of adoption, quality, productivity and business impact after release.

None of this is new. It is the sequence any operational change has always run on, and the ads describe it in ordinary language — licence reports, adoption at 90 days, overhead as a share of a leader's week.

What AI changes is the last step. Tools, models and costs keep moving after launch, so checking the number stops being a sign-off and becomes something that recurs. That makes the owner harder to skip, not easier.

4. The useful guardrails sit inside the work

The build-and-operate ads named concrete controls:

• a human check and confidence threshold;
• permission allowlists, sandboxing and approval gates;
• approved platforms and data-handling rules;
• output verification and a way to stop a high-risk pilot;
• monitoring for quality, cost and latency; and
• the judgement to leave a task human when a spreadsheet, script or phone call is better.

These controls serve an operating outcome. They are not a maturity ladder or a committee checklist. A lean business needs the same rigour as a large organisation, applied at a weight it can carry.

5. The rest of the August sample

• Who was advertising, across 45 roles. Five employers under 200 staff — four small businesses and one small public agency · two mid-sized businesses, one of them close enough to the line that it may really be a large one · 11 consultancies · one AI product company · 18 large organisations hiring directly · eight we did not size.
• AI was central in 38 roles and peripheral in 7. Every August role contained at least one AI duty; eight titles did not use the word “AI”.
• Thirteen roles disclosed a numeric salary or day rate. The permanent figures ranged from $80K–$110K to more than $195K; one contract paid $900–$1,000 per day. The roles are too different for a useful median.
• Named tools remained mixed. The records mention Microsoft Copilot in 13 distinct ads, Claude in 10, Azure in 10 and ChatGPT / OpenAI in 9. GitHub Copilot appears once and is counted separately — different product, different users.
• Five employers under 200 staff is a count, not a rate. Eight ads never showed us a company size, so five is the floor, not the answer. Three of the five were resolved after capture, which is why this edition's number is higher than June's or July's — the ads themselves got no smaller.
• July's operational-titles observation did not repeat. All five of August's under-200 employers advertised an explicit AI title. July's four filed the same kind of work under operational ones. The search frame changed between the two months, so this is a difference in what we asked for as much as a difference in what employers wrote.

The composite ad of the month

Composited from recurring August duties. No single employer is represented.

The role combines judgement, hands-on delivery, adoption, measurement and ongoing support. For a business with one AI budget and no spare headcount, that breadth is useful when it comes with enough seniority to make the risk/reward calls.

What this means if you run a small business

• Start with work you already understand. Internal quoting, reporting, reconciliation, onboarding and service workflows come with an owner and an operational number. That makes them easier to govern and measure than an isolated AI showcase.
• Name the owner before the tool. Someone needs to carry the work from decision through adoption and support. A vendor can help deliver it; the business still owns the outcome.
• Set the baseline before rollout. Use the measure you already run the business on: hours, cost, quality, conversion, tickets or rework. Check it again after people start using the change. A costed automation plan makes that comparison concrete.
• Use guardrails that earn their keep. Start with approved tools, access boundaries, a human check, output verification and an escalation path. Add more when the use case and risk justify it. The AI readiness assessment can help you take stock.
• If you hire, buy senior judgement and hands-on range. A lean business cannot split discovery, delivery, adoption and risk across five roles. One experienced operator can keep those decisions connected to the business outcome — which is what fractional technology leadership is for.

Cite this report

GraftPoint SME AI Hiring Index, August 2026 edition. Manual sample of 46 Melbourne-weighted AI and automation job-ad captures from 11–30 August 2026 (45 distinct roles). Of 32 programmes in the employer-adoption sample, 23 were internal only, 8 combined internal and customer-facing work, none were customer-facing only and 1 was unclear; the sample included planning, piloting, rollout and running stages. Methodology: graftpoint.com.au/sme-ai-hiring-index/methodology/