The 2027 Job Postings, Checked Three and a Half Months Later — Responsibility stayed with people, bridging jobs grew, and the entry route for young workers narrowed

A forecast is worth more for what its misses teach than for its hits. This piece sorts last time's forecasts into those that held, those that changed shape, and those that are still unknown. It uses only facts we could confirm on the source page.

Earlier piece: Reading the 2027 Job Listings in Advance (July 3, 2026)

01How we score — what counts as a hit

We use three grades. "Confirmed" means a fact directly shows the forecast direction. "Changed shape" means the direction was right but who does the work, or when, differed. "Still unknown" means there are no numbers good enough to judge yet.

One limit up front: job-postings counts for specific titles are mostly not public. We do not use figures that appear only in secondary write-ups or whose original page we could not open, so some widely shared numbers are absent here.

02Line one: auditing AI output — responsibility stayed with people, but with executives

Last time we forecast a job that audits AI-drafted work and takes responsibility for signing it off. That responsibility stays with people is confirmed in both regulation and practice.

Article 14 of the EU AI Act (the EU's AI regulation, adopted in 2024) requires high-risk AI to be designed so that it can be "effectively overseen by natural persons". In January 2026 the US Food and Drug Administration (FDA) and the European Medicines Agency (EMA) jointly issued guiding principles for AI in drug development. Principle 6 asks that data provenance, processing steps and analytical decisions be documented in a traceable and verifiable way. We wrote that designing audit trails would become work; regulators now ask for those trails in writing.

Audit firms are moving too. In April 2026 EY announced agentic AI (AI that carries out several steps without detailed human instructions) across its global assurance network, covering 130,000 assurance professionals and 160,000 audit engagements, while saying it will "maintain the fundamental role of human judgment, skepticism, and insight." Since June 2025 PwC has offered independent assurance of AI systems themselves, performed by teams of AI and machine-learning specialists.

Part of the forecast changed shape. In a KPMG survey of senior leaders at large US companies, fielded in May and June 2026 (technology-industry report), 61% placed accountability for AI-informed decisions with a named C-suite executive; 6% named a central AI governance committee. Responsibility stayed with people, but with executives rather than specialist auditors. In the same survey 67% cited human oversight skills as a challenge in deploying AI agents. People who can oversee are in short supply.

03Line two: supplying context — the work is discussed, the title is not yet confirmed

Last time we forecast that the work of turning company knowledge into a form AI can read would outlast the prompt engineer. We file this one under still unknown.

Atlan, which sells data-management software, wrote in March 2026: "Most organizations need context engineering work done. Fewer need a person whose entire job title says so." The need for the work is being discussed. But this time we found no public data counting postings for the title context engineer (someone who designs the background information given to AI).

Work comes first and the name follows, as we wrote last time. If that order holds, we are still before the name. Until postings can be counted by title, we cannot call this a hit or a miss.

04Line three: bridging — the strongest evidence so far

Last time we forecast three bridging jobs: designing how several AIs divide work, explaining AI decisions to regulators and auditors, and connecting AI with safety physicians in drug-safety monitoring. The first has the strongest evidence.

KPMG's third-quarter survey, released on September 24, 2026, found that 62% of organizations are building, deploying or developing AI agents, up from 53% the previous quarter. Organizations developing or implementing multi-agent systems rose to 25%, from 6% in each of the two prior quarters. Demand for people who design how several AIs split and hand back work rests on these numbers.

Postings figures appeared as well. According to Lightcast, the share of US job postings asking for agentic AI skills rose from 0.06% in 2024 to 0.23% in 2025, nearly 90,000 postings in the US alone. Indeed's figures for India show postings for forward-deployed engineers (engineers who work on the client's site to build AI into operations) grew nearly 19-fold between June 2025 and June 2026. Year-on-year growth then slowed to about 16 times in July and just over six times in August.

Here too the shape changed. We named an "AI orchestration" role; the title whose growth shows up in numbers is the engineer who goes on site and builds AI in.

Explaining AI to regulators has its legal basis confirmed, but its timing slipped. The EU's Digital Omnibus on AI entered into force on July 27, 2026. It moved obligations for high-risk uses (Annex III) from August 2, 2026 to December 2, 2027, and for AI in regulated products (Annex I) to August 2, 2028. That still falls within our 2027–2029 window, but the start of demand moved back by more than a year.

On drug-safety monitoring, a working group of the Council for International Organizations of Medical Sciences (CIOMS) published a report in December 2025. It warns that automation driven only by efficiency and cost targets may "risk eliminating the value added by humans in the loop." Meanwhile Jason Bryant of ArisGlobal, which sells safety-monitoring software, wrote in August 2026 that "the existence of human oversight alone does not in itself ensure reliable output." He describes focusing human review on low-confidence outputs, noting that regulators would still need to accept it. The job of connecting AI and safety physicians is shifting from checking everything to designing where to check.

05Credentials — private certifications came first; a national one appeared elsewhere

Last time we forecast that national credentials would not arrive by 2029 and that the path would run from in-house training to private certifications. That private certifications are multiplying first is confirmed.

ISACA, the information-systems audit association, launched Advanced in AI Audit (AAIA) in May 2025: an advanced, audit-specific AI certification for experienced auditors, open to holders of a CPA, CISA or CIA. ISACA has also announced an Advanced in AI Risk credential for the second quarter of 2026. Neither has published holder numbers.

Part of this changed shape. In January 2025 Taiwan's Ministry of Economic Affairs announced a national AI application planner certification exam. Our view that national credentials would come too late may not hold in every country. And the EU's Digital Omnibus weakened the AI literacy duty for organizations to "take measures supporting AI literacy." Regulation is pushing training and certification less than we assumed.

06A gap comes into view — boundary jobs need experience, and the entry route narrowed

Our earlier piece left something out: who gets into these boundary jobs, and how.

In June 2026 Stanford's Digital Economy Lab launched monthly indicators of the labor market built on anonymized ADP payroll data. According to Fortune, employment of workers aged 22 to 25 in highly AI-exposed occupations is shrinking at 3.8% a year. The data cover 4.6 million workers across more than 730 occupations.

Oversight, audit and bridging jobs all assume experience in making judgments. If KPMG's respondents name oversight skills as a barrier while young workers' employment falls, the places where people build that experience are shrinking. Last time we advised readers to start recording the reasons behind their judgments today. That advice still holds for individuals. For organizations there is a new task: deliberately keeping work where young people can build judgment.

Verification note: this piece uses only sentences and figures confirmed on the source pages. Figures found only in secondary write-ups, or on pages we could not open, are not used. Whether each fact appears in the sources was checked with Jev (TypeSafe's judgement model).