The World's Largest Staffing Firm Just Bet 50% of Its Revenue on AI Agents. Who's Maintaining the Documents They Read?
Adecco's unlimited Agentforce deal targets €12B in AI-driven revenue across 60 countries. The knowledge layer feeding those agents is the risk nobody's writing about.
The scale that makes this different
The Adecco Group, the world's largest staffing and talent services company, just signed an unlimited Agentforce 360 license with Salesforce. The stated target: AI agents driving over 50% of its revenues by the end of 2026 (PRNewswire).
At ~€24 billion in annual revenue, operating across 60+ countries, serving 100,000+ client companies and placing 500,000+ workers every year, that's roughly €12 billion in AI-agent-powered placements. The UK pilot showed 15% time savings for recruiters, reduced time-to-fill, and lower cost-to-serve. The numbers from the pilot are real.
Nobody's writing about what happens when the documents are wrong.
What those agents are actually reading
Every article covering the Adecco announcement focuses on the output layer: how fast agents process tasks, how many hours they save. The risk is upstream.
Staffing operations run on documents. When Adecco's agents go to work, they'll be reading five categories of them:
Job descriptions. Written once, updated rarely. An agent reading a JD that described requirements from six months ago places the wrong candidates. At 500,000 annual placements, "rarely" stops being rare fast.
Candidate profiles. Skills, availability, experience. Candidates update LinkedIn; they don't always update the staffing database. Stale profiles don't produce one bad match. They produce the same bad match, replicated across every agent that hits the same record.
Compliance documents. This is where it gets costly. Employment law changes constantly, and Adecco operates in 60+ countries. An agent making placement decisions in Germany, France, or Japan needs to read current compliance documentation. An outdated policy doesn't create an inefficiency; it creates an illegal placement.
Client contracts and requirements. Clients' needs change. The agent reading an 18-month-old contract doesn't know the client now wants Python engineers instead of Java developers. At volume, this is a recurring condition, not an exception.
Internal HR policies. Three global business units. 60 countries. Documents multiply and diverge. Agents reading conflicting policies produce inconsistent outcomes for the same transaction type.
Do the math. If agents drive €12 billion in revenue and even 15% of the underlying knowledge is stale or contradictory — that's €1.8 billion in placements where match quality is structurally degraded. Not because the AI failed. Because the document was wrong.
That's an illustrative estimate. It's not speculation; it's arithmetic.
The pattern playing out across the agentic wave
Adecco is the most quantified version of a problem that's been building for months.
Salesforce Data 360 (the data layer underpinning Agentforce) unifies structured CRM data across systems. It doesn't handle the unstructured document layer: the PDFs, policy archives, contracts, and compliance manuals that live outside CRM records. That gap is structural, and it's not unique to Adecco. The enterprise AI stack has a compute layer, a model layer, a data layer, and a knowledge layer. The first three are getting budgets. The fourth is being assumed.
Research on agentic AI failure rates points to the same upstream cause consistently: agents don't fail because models are bad. They fail because they retrieve against documents that were never built to be accurate at retrieval scale.
Then there's the Atlassian angle, which is hard to ignore this week. Atlassian cut 1,600 employees (10% of its workforce) citing AI as the driver, with most reductions from R&D. Atlassian builds Confluence, one of the most widely used enterprise knowledge bases in the world. The people who maintained and verified that knowledge are being cut. The AI agents depending on that knowledge are being scaled. Those two trends are moving in opposite directions and nobody's reconciling them.
Meanwhile, Foxit's research released this week found that enterprise executives report AI boosts productivity — but they net only 16 minutes per week in actual time savings, because 4 of every 4.6 perceived hours are consumed verifying AI outputs (Foxit/Sapio Research, March 2026). That verification burden is what happens when you can't trust the underlying documents.
The infrastructure gap nobody's announced
The Salesforce deal covers the agent layer. What it doesn't cover (because no CRM platform covers it) is the document knowledge layer beneath.
At Adecco's scale, that layer needs to do several things at once: ingest documents from multiple systems and formats across 60 countries (including scanned documents from legacy regional offices), detect contradictions across jurisdictions and business units before agents act on them, flag compliance documents that haven't been updated since a regulatory change, and propagate updates when employment law shifts. All without requiring manual re-uploading across three global business units and 60 country operations.
This is what knowledge management infrastructure does. Mojar AI works at exactly this layer: hybrid PDF parsing across document types and scan quality, contradiction detection across document sets, scheduled audits for document freshness, and conversational knowledge base updates through a management agent that doesn't require touching the underlying files manually. Salesforce Data 360 handles structured CRM records. The unstructured document knowledge layer is a different problem, and it needs different infrastructure.
Adecco isn't a Mojar customer. That's not the point. The point is that every enterprise scaling agentic AI revenue has the same gap, whether they've named it yet or not.
The question under the headline
The efficiency numbers from the UK pilot are real. The 50% revenue target is ambitious but grounded. CEO Denis Machuel's framing of human-centric AI implementation is coherent.
But Salesforce EVP Madhav Thattai put it directly: Adecco now has the "always-on foundation to connect millions of job seekers with career opportunities." That foundation only holds if it's accurate. An always-on retrieval system built on stale job descriptions, outdated compliance policies, and contradictory HR documents across 60 countries isn't a foundation. It's a liability that compounds at agent speed.
The question isn't whether AI agents can drive €12 billion in staffing revenue. They probably can. The question is whether the documents feeding those agents are trustworthy enough to put €12 billion behind them.
Most enterprises haven't asked that question yet. Adecco just made it very hard to keep ignoring.
Frequently Asked Questions
Adecco Group signed an unlimited Agentforce 360 license agreement with Salesforce in March 2026. The goal is for AI agents to drive more than 50% of Adecco's revenues — roughly €12 billion — by the end of 2026, across its three global business units operating in 60+ countries.
AI agents retrieve from documents to make decisions. When those documents — job descriptions, compliance policies, client contracts — are stale, contradictory, or never maintained, the agent's output is wrong at the same scale the agent operates. No model improvement fixes an upstream document accuracy problem.
Mojar AI operates at the document knowledge layer beneath platforms like Salesforce Agentforce. It ingests unstructured documents across formats (including scanned PDFs), detects contradictions across jurisdictions, flags documents that haven't been updated since a regulatory change, and enables conversational knowledge base updates without manual re-uploading.