Singapore

Proving PPC AI Skill Without a Certificate

As of November 2026, the most useful evidence of a PPC team's AI skill is a dated log of what the platform did, what the team accepted or overrode, and what happened to cost afterwards. A certificate cannot carry that record; account data can. When you size up an sem agency, ask which decisions they logged and what they measured on the other side.

That reading comes from two studies published in 2026, and it holds whether you are hiring or preparing your own review. The numbers below describe a general workforce, not paid search specifically; the transfer to an account is my own inference, and I say where it gets thin.

Key takeaways

72% of workers say their AI skills have never been formally evaluated, while 69% of managers call AI ability a major or moderate factor in hiring.
Only 20.6% of managers say their organisation can measure the ROI of AI spending, so the budget grows faster than the proof.
What a PPC team can show instead: a change log with reasons, a set of automated recommendations they rejected, and the performance attached to each.
A certificate records what someone knows about a platform's playbook. Neither study tests whether it predicts output on a live account.
Ask for a sample of accounts, a date range and the decision behind each override. Without the date range, a log is a story.

What the two surveys actually measured

Before any of these figures gets quoted in a review deck, the definition matters. CodeSignal commissioned two online surveys through Pollfish in September 2026, both samples entirely U.S.-based, and surveyed 1,000 U.S. employees and 1,000 U.S. people managers. The company defined knowledge workers by the work they do rather than by job title, and it did not screen out people who skip AI: 12% said they never use it at work.

The published report puts the total at 2,000 people managers and knowledge workers, which is the same pair of samples counted together. Treat the figure as two separate samples of 1,000, because a weighted average across them would not describe either group.

The headline pair: 72% of workers say their AI skills have never been formally evaluated, yet 69% of managers call AI ability a major or moderate factor in hiring. CodeSignal's own report side, drawn only from the manager sample, says 62.5% of managers made workforce decisions based on assumptions in the past 12 months. The two studies do not contradict each other; the first sums the manager and employee answers, the second reads the manager sample alone.

A third figure makes the pattern harder to dismiss. 86.4% of managers say candidates perform differently on the job than their resumes or interviews suggested. The resume became the proxy because nothing more direct was collected, and the survey describes what that costs when the proxy is wrong.

A manager compares a candidate's CV with live campaign dashboards during a hiring discussion.

Who is trusting what

Managers were asked where their skills data comes from. Manager observation or judgment led at 64.4%, employee self-assessments or surveys drew 47.1%, and certifications or credentials drew 45.7%. The first is subjective by construction; the third records course completion. Neither one is a sample of the work.

The gap shows up in the visibility figures. 41.7% of managers described their view of their teams' AI skills as very clear and 42.7% as somewhat clear. On the employee side, 71.6% said their AI-related skills had never been formally assessed or verified.

Two people on the same team can hold both beliefs without either lying. The manager sees the output, the employee knows what was never asked.

CodeSignal's list of barriers explains why the data stays thin: 16.9% of managers put it down to skills changing too quickly, 13.9% to data scattered across systems, and 12.7% to a lack of the right tools. Those are problems of plumbing as much as of attitude.

The training does not reach the work

62% of employees received employer training on how to use AI at work. Only 26% of the employees who had received employer-provided AI training said they could apply most or all of it in their daily work. 71.9% of employees said that AI training fitting their job or role was a major or minor barrier.

Where the skills do come from, the survey allowed multiple answers, so the shares do not add up to a round total. Free resources people found on their own led at 46%, trial and error on the job came next at 39%, and employer training trailed at 35%. Colleagues accounted for 22%, and self-paid courses for 12%.

Paid search has run on that last pair for years. Plenty of account managers learned bid strategy by breaking campaigns and reading practitioner posts at midnight. The difference now is who pays for the error. When the thing learned by trial and error is Google's own automation, each lesson spends a client's budget, and the record of that lesson usually lives in someone's memory.

For a team deciding what to ask of an agency, this is the useful part of the survey: the skill was built outside the training programme, so the proof will not be found in the training certificate either.

What the budget is buying

84.2% of managers said their organisation's investment in AI tools and capabilities had increased over the past 12 months. Asked what their organisation could measure effectively today, only 20.6% of managers selected ROI from AI spending.

That single pair carries the case for a decision log. An account team that cannot attach a result to an automated change is in the same position as the managers in the survey: spend visible, return asserted.

The honest caveat is that "measure the ROI of AI spending" in the survey means a workforce and tooling decision, not a campaign. The mechanics still transfer, because both need the same thing: a change, a date, and a measured outcome on the other side of it.

Question the survey leaves open	What closes it in a Google Ads account
Was any skill ever assessed?	A list of automated recommendations the team rejected, with reasons
Is the manager's view accurate?	The same accounts read by two people against the same checklist
Did training change the work?	Account samples before and after the training date
Can the spend be justified?	Cost per conversion or ROAS attached to each logged change

Building an override log you can defend

The routine below is my own, built on the survey's finding that judgement has to be visible in the work. It needs no new software; the platform already keeps the raw material.

Fix the definition and the window. Pick the metric before you start, and state the comparison period. A log that opens with "ROAS improved" leaves the reader to guess whether the window is a week or a quarter.
Record the decision, not just the change. A sem google ads account's change history will tell you that a campaign moved to a different bid strategy. It will not tell you why, or what you checked before the move.
Attach the result after the fact. When the window closes, write the outcome next to the decision, including the ones that went wrong. A log with no losses reads as a marketing document.
Keep rejected recommendations. Most teams only record what they did. What automation proposed and a person refused is where the judgement shows.
Sample a few accounts before and after training. The survey's own advice points this way, and it answers the question a manager actually has.

For a smaller team with no assessment budget, the survey's suggestion of starting with managers translates into one step: write down what good AI use looks like for each role, in terms of account work. For a specialist, that might be catching irrelevant search terms inside a stated number of days. For a strategist, it might be deciding whether conversion tracking is clean enough to justify value-based bidding.

Reading an override log without being fooled

A decision log is evidence only if it survives three questions, and the first one catches most bad logs.

Who supplied the denominator? A claim that "nine of ten overrides worked" means little unless the ten were all the overrides made. A log that lists only successful overrides is a sample chosen after the result, and the selection effect alone can produce a flattering number.

Is the window the same for every decision? Bid strategy changes need days to settle, and conversion lag stretches that. Comparing a change measured over a week against one measured over a month is a definition error, not a strong result.

Does the log survive a second reader? Ask someone outside the account team to trace one decision from the log to the change history and to the performance report. If the three do not line up, the log describes an intention, not a record.

A reviewer traces a logged decision from a printed log to a screen and a performance report.

What a certificate does and does not show

Google's AI-Powered Performance Ads Certification on Skillshop is free, and third-party study guides describe it as a timed, 50-question exam. It is worth taking, because it teaches how Google recommends its AI campaign types be used. It records knowledge of a playbook.

The gap is structural. The exam is written by the platform whose recommendations a good team sometimes overrides, and it cannot show the afternoon someone declined to follow them, or what cost per acquisition did afterwards. A Skillshop badge is a completion metric, and no client should judge a campaign on completion alone.

For a buyer, the two evidence types answer different questions. Ask for the certificate when you want to know what the team has been taught. Ask for the log when you want to know what they have decided. On a sem agency shortlist, ask for both and read the log first, because it is the one that can be checked.

What Topkee's own systems log

The survey's finding is that judgement has to be visible in the work, and that is the design point to check a vendor's system against. Topkee runs paid search as an official Google Ads partner, and its Premier Partner standing dates from a July award in 2016; the same governance shows up in how its TTO Funnel Tuning Agent reports.

Under Topkee's product notes, every action the agent proposes starts as pending confirmation. Before authorisation it audits, prioritises and records each action with its reason, and each one lands in one of four states: auto-execute, route to Topkee experts, hold and observe, or archive. The stated delivery commitment is zero unauthorised account changes.

Each recommendation is graded P0, P1 or P2 and tagged with its target level, object ID and the specific data basis, across seven levels from ad to click to page to enquiry to qualified lead to quote to closed deal. Reports carry a four-step validation logic of metric, judgement, weighting and action, and the evidence is filed as documents rather than described in conversation. Each report also states the next cycle's expected range and its dependencies, which the following cycle checks automatically. That is the level of record a buyer should hold any sem agency to, whatever tools it runs.

None of this is an AI skill assessment, and the notes do not claim it is. It is a run record, and a run record is not by itself independent verification of completion. What it gives a client is the raw material for the questions above: a decision, its reason, and something to check it against.

What to ask before you renew

Five requests, phrased so that any team can answer them or admit they cannot.

A dated sample of automated recommendations they rejected, with the reason and the result.
The metric definition and the comparison window used for each logged decision.
One decision traced end to end by a person outside the account team.
The account health review they ran at the start, including the parts they found hard to measure.
The next cycle's expected range, written down before the cycle starts.

A sem google ads team that keeps a readable change log will answer most of these in an afternoon. The ones it cannot answer tell you what the review is missing.

Where a team can produce these, the AI skill question mostly answers itself, because the work is on the table. Where it cannot, the survey's 62.5% figure describes the alternative: a decision made on how confident the pitch sounded.

Topkee's free tracking and funnel audit documents what is measurable and what is not before any spend moves, which is a reasonable place to start if you want the "before" state recorded. For teams that want the account-level view of where AI bidding is helping and where it leaks, the report on what Google's Q3 results mean for your SEM budget is a useful companion piece, and how brands are lifting ROI with AI-powered Google Ads shows the demand-generation side of the same question.

A client and account team review dated printouts and an audit folder at a table before renewal.

FAQ

Does a Skillshop certificate prove a PPC manager uses AI well?

It proves they know how Google recommends its AI campaign types be used, and third-party guides describe it as a timed, 50-question exam. It does not record what they overrode on a live account or what happened afterwards. Use it as evidence of training, not of judgement.

What is an override log in Google Ads?

A record of each automated recommendation the team acted on, written with the reason, the date and the result. Change history supplies the "what" and the platform does not supply the "why", so the reason has to be written separately.

How many accounts should a client sample?

Enough to see whether the behaviour repeats. A sample of a few accounts per person is the survey's own advice, and it matters more that the sample is chosen before the results are known than that it is large.

Can a team measure ROI on AI-driven bidding?

Cost per conversion and ROAS can be attached to a logged change with a fixed window. The survey's 20.6% figure sits at the level of workforce tooling, where ROI is harder to isolate. Do not read it as a benchmark for a single Google Ads account.

Is a run record the same as a verified result?

No. A run record ties the task list to the evidence behind each step, which is what makes a decision traceable. It is not independent verification that the outcome followed from the decision, and any team presenting it as such is overreaching.