Salesforce Agent Evidence Readiness: An 11 Layer Check Before Dreamforce · Enterprise Agentic AI Insights

Salesforce names eleven agentic architecture layers. Two carry the audit evidence and nobody owns them. Run the two minute check before Dreamforce.

Six days from now, roughly forty thousand people walk into Moscone Center for Dreamforce. More than 1,600 breakout sessions. Fifty-plus keynotes. The published spine of the whole event is the Agentic Enterprise, and governance at scale is named as a pillar.

Every one of them comes home with a scaling plan.

Almost none come home with an evidence plan.

That is the gap this piece is about, and there is now a number on it that did not exist last week.

What the CFO survey actually said

Salesforce published its CFO Priorities Report this morning. A double-blind survey of 865 finance leaders across five countries, fielded in May. Respondents did not know Salesforce commissioned it.

Nearly three quarters say the scope of their role grew in the last year. The top-cited reason, ahead of revenue complexity, ahead of the growing tech stack, ahead of everything else they were asked about, is managing AI use and expansion. Half now describe themselves as primary decision makers in their company's AI strategy.

Then the part worth reading twice. Here is what those same leaders name as the blockers to expanding AI.

<figure class="ad-fig" style="margin:30px 0"> <svg viewBox="0 0 720 300" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="Bar chart of the top five blockers to expanding AI use named by 865 finance leaders. Security 46 percent. Limited internal subject matter expertise 42 percent. Governance and compliance concerns 42 percent. Integration with existing stack 42 percent. Lack of funding or resources 39 percent."> <rect width="720" height="300" fill="#0D0F12"/> <text x="24" y="30" fill="#EDE8E3" font-family="Roboto,Arial" font-size="17" font-weight="700">What is actually blocking AI expansion</text> <text x="24" y="52" fill="#07A3D1" font-family="Roboto,Arial" font-size="13" font-weight="600">865 finance leaders, Salesforce CFO Priorities Report, September 9 2026</text> <g font-family="Roboto,Arial" font-size="13" font-weight="600"> <rect x="330" y="74" width="322" height="26" rx="4" fill="#FF4700"/><text x="24" y="92" fill="#EDE8E3">Security concerns</text><text x="662" y="92" fill="#FF8A5C" font-weight="700">46%</text> <rect x="330" y="110" width="294" height="26" rx="4" fill="#FF7900"/><text x="24" y="128" fill="#EDE8E3">Limited internal expertise</text><text x="662" y="128" fill="#FFA340" font-weight="700">42%</text> <rect x="330" y="146" width="294" height="26" rx="4" fill="#FF7900"/><text x="24" y="164" fill="#EDE8E3">Governance and compliance</text><text x="662" y="164" fill="#FFA340" font-weight="700">42%</text> <rect x="330" y="182" width="294" height="26" rx="4" fill="#07A3D1"/><text x="24" y="200" fill="#EDE8E3">Integration with the stack</text><text x="662" y="200" fill="#5CC6EA" font-weight="700">42%</text> <rect x="330" y="218" width="273" height="26" rx="4" fill="#07A3D1" opacity="0.72"/><text x="24" y="236" fill="#EDE8E3">Funding and resources</text><text x="662" y="236" fill="#5CC6EA" font-weight="700">39%</text> </g> <line x1="330" y1="68" x2="330" y2="252" stroke="#EDE8E3" stroke-opacity="0.25"/> <text x="24" y="278" fill="#EDE8E3" fill-opacity="0.72" font-family="Roboto,Arial" font-size="12">Three of the top four are control problems, not model problems.</text> </svg> <figcaption style="font-family:Roboto,Arial,sans-serif;font-size:13px;line-height:1.5;color:rgba(237,232,227,.72);margin-top:10px">Figure 1. Primary blockers to expanding AI use. Salesforce CFO Priorities Report, 865 finance leaders, published September 9, 2026.</figcaption> </figure>

Security at 46 percent. Governance and compliance at 42. Integration at 42. And sitting at exactly the same 42 percent, limited internal subject matter expertise.

Three of the top four are control problems. None of them are model problems. Nobody in this survey said the models were not good enough.

Sam Chung, who runs customer for Agentforce Revenue Management, put the shift in one line:

> "AI decisions don't reach finance as budget requests anymore. They reach us as questions about risk, control, and accountability. The CFO has moved from signing off on AI to being answerable for it."

Read that as a CIO and it becomes an operational instruction. Somebody in your building is now personally answerable for agent behavior, and the evidence they will be answerable with is evidence you have to produce.

The mechanic that is failing

Here is the specific thing that breaks, and it is not a Salesforce problem. It is an every-platform problem that happens to be legible in Salesforce because Salesforce published its architecture.

Salesforce's Architecture Center describes the agentic enterprise as eleven layers. Nine stack vertically. Two run across all nine.

<figure class="ad-fig" style="margin:30px 0"> <svg viewBox="0 0 720 320" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="Diagram of the eleven layer Salesforce agentic architecture. Two cross layers, Security and Governance and IT Operations and Observability, span all nine stacked layers. The four agentic layers are Agentic, Semantic, AI and ML, and Enterprise Orchestration. The five traditional layers are Experience, Applications and App Services, Data, Integration, and Infrastructure."> <rect width="720" height="320" fill="#0D0F12"/> <text x="24" y="28" fill="#EDE8E3" font-family="Roboto,Arial" font-size="17" font-weight="700">Eleven layers. Two of them belong to everyone.</text> <g font-family="Roboto,Arial" font-size="13" font-weight="600"> <rect x="24" y="46" width="86" height="240" rx="8" fill="#FF4700" fill-opacity="0.20" stroke="#FF4700" stroke-width="2"/> <text x="67" y="150" fill="#FF8A5C" font-size="12" text-anchor="middle">SECURITY</text> <text x="67" y="168" fill="#FF8A5C" font-size="12" text-anchor="middle">AND</text> <text x="67" y="186" fill="#FF8A5C" font-size="12" text-anchor="middle">GOVERNANCE</text> <rect x="610" y="46" width="86" height="240" rx="8" fill="#FF4700" fill-opacity="0.20" stroke="#FF4700" stroke-width="2"/> <text x="653" y="150" fill="#FF8A5C" font-size="12" text-anchor="middle">IT OPS AND</text> <text x="653" y="168" fill="#FF8A5C" font-size="12" text-anchor="middle">OBSERV-</text> <text x="653" y="186" fill="#FF8A5C" font-size="12" text-anchor="middle">ABILITY</text> <rect x="124" y="46" width="472" height="24" rx="4" fill="#FF7900" fill-opacity="0.34" stroke="#FF7900"/><text x="360" y="63" fill="#EDE8E3" text-anchor="middle">Agentic Layer</text> <rect x="124" y="74" width="472" height="24" rx="4" fill="#FF7900" fill-opacity="0.34" stroke="#FF7900"/><text x="360" y="91" fill="#EDE8E3" text-anchor="middle">Semantic Layer</text> <rect x="124" y="102" width="472" height="24" rx="4" fill="#FF7900" fill-opacity="0.34" stroke="#FF7900"/><text x="360" y="119" fill="#EDE8E3" text-anchor="middle">AI and ML Layer</text> <rect x="124" y="130" width="472" height="24" rx="4" fill="#FF7900" fill-opacity="0.34" stroke="#FF7900"/><text x="360" y="147" fill="#EDE8E3" text-anchor="middle">Enterprise Orchestration</text> <rect x="124" y="162" width="472" height="24" rx="4" fill="#07A3D1" fill-opacity="0.28" stroke="#07A3D1"/><text x="360" y="179" fill="#EDE8E3" text-anchor="middle">Experience</text> <rect x="124" y="190" width="472" height="24" rx="4" fill="#07A3D1" fill-opacity="0.28" stroke="#07A3D1"/><text x="360" y="207" fill="#EDE8E3" text-anchor="middle">Applications and App Services</text> <rect x="124" y="218" width="472" height="24" rx="4" fill="#07A3D1" fill-opacity="0.28" stroke="#07A3D1"/><text x="360" y="235" fill="#EDE8E3" text-anchor="middle">Data</text> <rect x="124" y="246" width="472" height="24" rx="4" fill="#07A3D1" fill-opacity="0.28" stroke="#07A3D1"/><text x="360" y="263" fill="#EDE8E3" text-anchor="middle">Integration</text> <rect x="124" y="274" width="472" height="12" rx="4" fill="#07A3D1" fill-opacity="0.28" stroke="#07A3D1"/> </g> <text x="124" y="284" fill="#EDE8E3" font-family="Roboto,Arial" font-size="12" font-weight="600" text-anchor="start">Infrastructure</text> <text x="24" y="308" fill="#EDE8E3" fill-opacity="0.72" font-family="Roboto,Arial" font-size="12">Orange, the four new agentic layers. Teal, the five traditional layers, re-engineered. Red, the two cross layers.</text> </svg> <figcaption style="font-family:Roboto,Arial,sans-serif;font-size:13px;line-height:1.5;color:rgba(237,232,227,.72);margin-top:10px">Figure 2. The eleven layer agentic architecture. The two cross layers span all nine and are the ones that carry the evidence. Source: Salesforce Architecture Center.</figcaption> </figure>

The nine stacked layers each have an owner. Somebody's name is on Data. Somebody's name is on Integration. There is a team, a budget line, a backlog.

The two cross layers are Security and Governance, and IT Operations and Observability. They belong to everyone, which in practice means they appear on no one's ticket. They are the two layers that carry every question an examiner will ask, and they are the two that get funded last because no single team owns the shortfall.

That is the mechanic. Not agent sprawl. Agent sprawl is the symptom. The cause is that the two layers holding the evidence are structurally unowned.

You can watch it happen in the defaults. The Einstein Trust Layer audit trail is off by default and retains roughly ninety days. Agentforce Session Tracing is a separate toggle. Event Monitoring exports, but somewhere off platform, which is a project rather than a checkbox. Every one of those is a deliberate, reasonable product decision. Together they mean that the estate which has not turned them on cannot reconstruct what its agents did last quarter, and cannot know that until somebody asks.

Why it stops being an IT problem

This is the turn, and it is worth being precise about it rather than dramatic.

EU AI Act. Since August 2, 2026, the AI Office and Member State authorities hold enforcement powers. They can demand documentation, run evaluations, require corrective measures, order withdrawal, and fine. For general-purpose model obligations the ceiling is 15 million euro or 3 percent of worldwide annual turnover, whichever is higher. Article 50 transparency obligations are applicable now, and they land squarely on the Experience layer: a person dealing with an agent has to know that is what they are dealing with.

A note on what we are not saying. Some commentary this month claims technical file audits for high-risk systems begin in September. The Commission's digital omnibus defers high-risk applicability, with Annex III moving to December 2027 and Annex I to August 2028. Where market commentary and primary legal sources disagree, we follow the primary sources. Transparency obligations are live. High-risk classification is not, yet.

United States. OCC Bulletin 2026-13, issued April 17, 2026, leaves agentic systems outside the federal model risk perimeter. Read carefully, that is not permission. It means no regulator has written your standard for you, so your own second line has to, and will be held to whatever it writes.

Put those together with the survey and the position is uncomfortable in a specific way. The person now answerable for AI risk is in finance. The evidence sits in an IT layer nobody owns. And the regulator has either just been handed enforcement powers, or has declined to define the standard at all.

The move

The useful response is not a governance program. It is one page.

For each of the eleven layers, answer three questions. What evidence would an examiner ask for here. Where does our estate produce that evidence today. And could we hand it over this week without building anything first.

That third question is the whole exercise. Configured is not evidenced. Most estates have human intervention configured through Flow approvals and Agentforce write confirmations. Very few retain a record of where a human actually intervened, which is the thing an examiner asks for.

Try it now. Eleven layers, one honest answer each.

<div id="ad-xwalk" class="adx"> <style> .adx{--o:#FF4700;--a:#FF7900;--c:#07A3D1;--bg:#0D0F12;--w:237,232,227;font-family:Roboto,"Helvetica Neue",Arial,sans-serif;background:var(--bg);color:rgba(var(--w),.88);border-radius:18px;padding:28px;margin:34px 0;border:1px solid rgba(var(--w),.14);position:relative;overflow:hidden} .adx *{box-sizing:border-box} .adx:before{content:"";position:absolute;inset:0 0 auto;height:5px;background:linear-gradient(90deg,var(--o),var(--a) 45%,var(--c))} .adx h3{font-size:27px;line-height:1.2;margin:8px 0 6px;color:rgba(var(--w),.98);font-weight:900} .adx .eyebrow{font-size:13px;letter-spacing:.13em;text-transform:uppercase;color:var(--c);font-weight:700} .adx .lede{font-size:16px;line-height:1.55;color:rgba(var(--w),.8);margin:0 0 18px;max-width:64ch} .adx .row{display:flex;gap:14px;align-items:flex-start;padding:13px 14px;border-radius:11px;background:rgba(var(--w),.05);margin-bottom:9px;border:1px solid rgba(var(--w),.09);transition:.15s} .adx .row:hover{background:rgba(var(--w),.08)} .adx .row.on{border-color:rgba(7,163,209,.6);background:rgba(7,163,209,.11)} .adx .row input{margin-top:4px;width:19px;height:19px;accent-color:#07A3D1;flex:none;cursor:pointer} .adx .row label{cursor:pointer;flex:1} .adx .ttl{font-size:16px;font-weight:700;color:rgba(var(--w),.95);display:block;margin-bottom:3px} .adx .ev{font-size:14px;line-height:1.5;color:rgba(var(--w),.76);display:block} .adx .src{font-size:13px;line-height:1.45;color:rgba(var(--w),.62);display:block;margin-top:5px} .adx .tag{display:inline-block;font-size:12px;font-weight:700;letter-spacing:.05em;padding:2px 8px;border-radius:5px;margin-right:7px;vertical-align:2px} .adx .t-x{background:rgba(255,71,0,.2);color:#FF8A5C;border:1px solid rgba(255,71,0,.55)} .adx .t-a{background:rgba(255,121,0,.18);color:#FFA340;border:1px solid rgba(255,121,0,.5)} .adx .t-t{background:rgba(7,163,209,.18);color:#5CC6EA;border:1px solid rgba(7,163,209,.5)} .adx .grp{font-size:13px;font-weight:700;letter-spacing:.11em;text-transform:uppercase;color:rgba(var(--w),.66);margin:20px 0 9px} .adx .out{margin-top:20px;padding:20px;border-radius:14px;background:linear-gradient(135deg,rgba(255,71,0,.13),rgba(7,163,209,.13));border:1px solid rgba(var(--w),.2)} .adx .lvl{font-size:33px;font-weight:900;line-height:1.15;margin:0 0 8px} .adx .verdict{font-size:16px;line-height:1.55;margin:0 0 14px;color:rgba(var(--w),.88)} .adx .bar{height:12px;border-radius:6px;background:rgba(var(--w),.13);overflow:hidden;margin:12px 0 6px} .adx .fill{height:100%;background:linear-gradient(90deg,var(--o),var(--a),var(--c));transition:width .4s} .adx .scale{display:flex;justify-content:space-between;font-size:13px;color:rgba(var(--w),.68);font-weight:600} .adx .cta{display:inline-block;margin-top:14px;background:var(--o);color:#fff;font-size:15px;font-weight:700;padding:12px 22px;border-radius:9px;text-decoration:none} .adx .cta.alt{background:transparent;border:2px solid var(--c);color:#5CC6EA;margin-left:9px} .adx .foot{font-size:13px;color:rgba(var(--w),.6);margin-top:16px;line-height:1.5} @media(max-width:600px){.adx{padding:20px 16px}.adx h3{font-size:23px}.adx .lvl{font-size:27px}.adx .cta.alt{margin:9px 0 0}} </style> <div class="eyebrow">Interactive · 2 minutes</div> <h3>The Agentic Evidence Check: 11 layers, one honest answer each</h3> <p class="lede">Salesforce publishes an eleven layer architecture for the agentic enterprise. For each layer, tick the box only if you could hand an examiner the evidence this week without building anything first. Nothing is sent anywhere. The score runs in your browser.</p>

<div class="grp">Cross layers, the two nobody owns</div> <div id="adx-rows"></div>

<div class="out"> <div class="lvl" id="adx-lvl">Tick the layers you can evidence</div> <p class="verdict" id="adx-verdict">Your result plots two points on Salesforce's own L1 to L4 Agentic Maturity Model: the level your agents behave at, and the level your controls can prove. The distance between them is the finding.</p> <div class="bar"><div class="fill" id="adx-fill" style="width:0%"></div></div> <div class="scale"><span>L1 Assisted</span><span>L2 Guided</span><span>L3 Delegated</span><span>L4 Autonomous</span></div> <div id="adx-gaps" style="margin-top:14px"></div> <a class="cta" href="https://vufubcgxqbebbvrqizjl.supabase.co/storage/v1/object/public/insights-assets/insights/2026-09-09/aegis-salesforce-crosswalk.pdf" download="AEGIS-Salesforce-agentic-crosswalk.pdf">Get the full crosswalk, one page</a> <a class="cta alt" href="https://ariana.digital/ai-governance.html">See the AEGIS framework</a> <p class="foot">Sources: Salesforce Architecture Center, The Agentic Enterprise IT Architecture · Salesforce CFO Priorities Report, September 9, 2026 · EU AI Act Article 50, applicable since August 2, 2026 · OCC Bulletin 2026-13. Not legal advice.</p> </div> </div> <script> (function(){ var D=[ {g:"Cross layers, the two nobody owns",t:"Security and Governance",k:"x",e:"Permission derivation for every agent action, scoped to task intent rather than inherited wholesale.",s:"Profiles, permission sets, sharing rules, field level security. Static and role based, which is the starting point the architecture says to move off. Trusted Agent Identity in Agent Fabric is the upgrade path."}, {t:"IT Operations and Observability",k:"x",e:"An immutable, time bounded record of each agent step, retained past your audit cycle.",s:"Einstein Trust Layer audit trail, off by default and roughly 90 days. Agentforce Session Tracing, separate toggle. Event Monitoring, exportable. Retention and immutability both need an off platform destination."}, {g:"The four agentic layers",t:"Agentic Layer",k:"a",e:"A current inventory of agents, their versions, permitted tools, and who approved each one.",s:"Agent Builder, agent and tool registries. Agent Fabric adds Agent Scanners and Controlled Registration, which reach Bedrock and Microsoft Foundry. Most gaps are agents arriving from outside the org."}, {t:"Semantic Layer",k:"a",e:"What the agent could reason over, and whether the definitions it used were the approved ones.",s:"Data 360, catalog and metadata services. Enterprise Knowledge Graph maturity is where most estates are thinnest."}, {t:"AI and ML Layer",k:"a",e:"Which model produced which output, under what guardrail, at what version.",s:"Einstein Trust Layer masking and scoring, model gateway behavior. LLM Governance on AI Gateway centralizes this across a multi model stack. OCC Bulletin 2026-13 leaves agentic outside the federal model risk perimeter, so your second line writes this standard."}, {t:"Enterprise Orchestration",k:"a",e:"Proof that a human could intervene, and a record of where one actually did.",s:"Flow, approval steps, Agentforce write confirmations. Trusted Agent Identity adds mobile approval for high stakes actions such as money movement. Intervention is usually configured and rarely evidenced."}, {g:"The five traditional layers, re-engineered",t:"Experience",k:"t",e:"That a person knew they were dealing with an agent, and could see the basis for its answer.",s:"Attribution and transparency UI patterns. This is the row EU AI Act Article 50 touches directly, applicable since August 2, 2026."}, {t:"Applications and App Services",k:"t",e:"Which actions were exposed to agents, with what guardrails and rate limits.",s:"Modular services, app side guardrails, invocable actions. MCP Bridge adds rate limiting to existing APIs without code changes."}, {t:"Data",k:"t",e:"Where the data came from, where it went, and whether it left the permitted region.",s:"Data 360 lineage, Hyperforce residency."}, {t:"Integration",k:"t",e:"How the agent reached the estate, and whether that path was governed.",s:"MCP server and Event Monitoring, API gateway. This is the row Claudeforce sits in when Salesforce in Claude reaches open beta."}, {t:"Infrastructure",k:"t",e:"That agent workload cannot exhaust capacity or budget without a signal firing first.",s:"Flex Credits consumption, pre-commit structures. LLM Governance on AI Gateway adds central cost control. No native monthly spend ceiling."} ]; var TAG={x:["Cross layer","t-x"],a:["Agentic","t-a"],t:["Traditional","t-t"]}; var host=document.getElementById("adx-rows");if(!host)return; var html=""; D.forEach(function(r,i){ if(r.g&&i>0){html+='<div class="grp">'+r.g+'</div>';} html+='<div class="row" id="adxr'+i+'"><input type="checkbox" id="adxc'+i+'" data-i="'+i+'"><label for="adxc'+i+'">'+ '<span class="ttl"><span class="tag '+TAG[r.k][1]+'">'+TAG[r.k][0]+'</span>'+r.t+'</span>'+ '<span class="ev">Evidence asked for: '+r.e+'</span>'+ '<span class="src">Where a Salesforce estate produces it: '+r.s+'</span></label></div>'; }); host.innerHTML=html; function score(){ var n=0,gaps=[]; D.forEach(function(r,i){ var c=document.getElementById("adxc"+i),row=document.getElementById("adxr"+i); if(c.checked){n++;row.classList.add("on");}else{row.classList.remove("on");gaps.push(r.t);} }); var pct=Math.round(n/11*100); document.getElementById("adx-fill").style.width=pct+"%"; var lvl,col,verdict; if(n===0){lvl="Tick the layers you can evidence";col="rgba(237,232,227,.9)";verdict="Your result plots two points on Salesforce's own L1 to L4 Agentic Maturity Model: the level your agents behave at, and the level your controls can prove. The distance between them is the finding.";} else if(n<=3){lvl="Evidenced at Level 1";col="#FF4700";verdict="Eight or more layers you cannot evidence this week. If any agent in your estate takes an action without a person in the loop, you are running Level 2 or higher behavior on Level 1 controls. That sentence belongs in front of your risk committee before it belongs in a budget request.";} else if(n<=6){lvl="Evidenced at Level 1 to 2";col="#FF7900";verdict="You can evidence the perimeter but not the reasoning. This is the most common position, and it is the one that fails on the second question in a security review rather than the first. Close the observability and orchestration rows before you widen the rollout.";} else if(n<=9){lvl="Evidenced at Level 2 to 3";col="#07A3D1";verdict="Genuinely ahead of the field. The remaining gaps are usually retention, immutability, and evidencing human intervention rather than configuring it. Those are the three an examiner opens with.";} else{lvl="Evidenced at Level 3 to 4";col="#5CC6EA";verdict="You can reconstruct agent behavior end to end. Worth pressure testing the claim against a live incident replay rather than a control inventory, because the two rarely agree the first time.";} var L=document.getElementById("adx-lvl");L.textContent=lvl;L.style.color=col; document.getElementById("adx-verdict").textContent=verdict; var g=document.getElementById("adx-gaps"); if(n>0&&gaps.length){g.innerHTML='<div style="font-size:14px;line-height:1.6;color:rgba(237,232,227,.82)"><strong style="color:#FF8A5C">Layers you cannot evidence this week ('+gaps.length+'):</strong> '+gaps.join(" · ")+'</div>';} else if(n===11){g.innerHTML='<div style="font-size:14px;color:rgba(237,232,227,.82)"><strong style="color:#5CC6EA">All eleven layers evidenced.</strong> Take this into your next incident replay and see if it holds.</div>';} else{g.innerHTML="";} } host.addEventListener("change",score);score(); })(); </script>

What to do with your number

If you could not tick eight or more, you are almost certainly running Level 2 agent behavior on Level 1 controls, in the language of Salesforce's own Agentic Maturity Model. That is a sentence with a number in it, which is the kind of sentence that survives a risk committee. A heat map is not.

It also gives you a specific ask for next week. Instead of walking the Dreamforce floor asking what else agents can do, walk it asking four questions:

Retention and immutability. Where does our agent audit trail go after ninety days, and who can alter it before then?
Inventory completeness. Does our agent inventory include agents that reach the org from outside it? Agent Fabric's Agent Scanners now reach Bedrock and Microsoft Foundry, and Controlled Registration only admits agents that meet your rules. Both are generally available, both licensed through MuleSoft, and both off in most estates.
Evidenced intervention. Trusted Agent Identity lets an agent act under a specific user's permissions and pushes a mobile approval for high-stakes actions such as money movement. Is that switched on for the actions where it matters, and is the approval retained?
Spend ceiling. There is no native monthly cap. LLM Governance on the AI Gateway centralizes cost control across a multi-model stack. Who is watching consumption between now and the renewal?

None of those four are things you will find on the session catalog. All four are things the vendor will answer honestly if you ask in the room.

The part the survey buried

Go back to the blockers chart and look at the one people skip. Limited internal subject matter expertise, 42 percent, tied with governance.

The finance function has been handed accountability for AI risk and is telling a double-blind survey it does not have the people to carry it. That is not a tooling gap and no platform purchase closes it. It closes with principal led, domain-savvy AI-ready teams who have done this inside a regulated estate before, and who can tell the difference between a control that is configured and a control that is evidenced.

Six days. The scaling plan will write itself next week. Bring the evidence plan with you.

---

Sources

Salesforce, "New Research: CFOs Turn to AI and Agents to Tame Growing Revenue Complexity," September 9, 2026 · https://www.salesforce.com/news/stories/cfos-ai-agents-to-tame-revenue-complexity/
Salesforce, CFO Priorities Report, methodology: double-blind survey, 865 finance leaders, France, Germany, Japan, United Kingdom, United States, fielded May 4 to 15, 2026 · https://www.salesforce.com/sales/resources/cfo-priorities-report/
Salesforce Architecture Center, "The Agentic Enterprise: The IT Architecture for the AI-Powered Future" · https://architect.salesforce.com/fundamentals/agentic-enterprise-it-architecture
Salesforce, "Salesforce Advances Agent Fabric: New Guided Determinism and Governance Controls to Scale Multi-Vendor AI Faster," April 15, 2026 · https://www.salesforce.com/news/stories/agent-fabric-control-plane-announcement/
MuleSoft, Trusted Agent Identity · https://www.mulesoft.com/ai/trusted-agent-identity/announcement
MuleSoft, AI Gateway · https://www.mulesoft.com/platform/ai-gateway
European Commission, "Commission starts enforcing AI Act rules and new transparency requirements on 2 August" · https://digital-strategy.ec.europa.eu/en/news/commission-starts-enforcing-ai-act-rules-and-new-transparency-requirements-2-august
European Commission, The enforcement framework of the AI Act · https://digital-strategy.ec.europa.eu/en/policies/enforcement-ai-act
Wilson Sonsini, "EU AI Act Enforcement Phase Begins" · https://www.wsgr.com/en/insights/eu-ai-act-enforcement-phase-begins.html
Office of the Comptroller of the Currency, Bulletin 2026-13, April 17, 2026 · https://www.occ.gov/news-issuances/bulletins/2026/bulletin-2026-13.html
Salesforce, "Salesforce and Anthropic Announce Claudeforce," August 26, 2026 · https://www.salesforce.com/news/press-releases/2026/08/26/salesforce-and-anthropic-announce-claudeforce/
MarketScale, "Dreamforce 2026 puts the agentic enterprise on trial in San Francisco this September" · https://www.marketscale.com/industries/software-and-technology/dreamforce-2026-puts-the-agentic-enterprise-on-trial-in-san-francisco-this-september
Ariana Digital, AI governance and the AEGIS framework · https://ariana.digital/ai-governance.html

---

Method and correction policy

Every edition is researched fresh against sources published within the preceding seven days where the item is time-sensitive. Figures carry a chip: VERIFIED means named, dated and publicly checkable; CITED means named source, not independently re-verified; FLAG means contested and pending re-verification. Where market commentary conflicted with primary legal sources this week, notably on EU high-risk applicability, we followed the primary legal sources and said so.

Open the formatted article on Ariana.Digital →