# Automation and Technology's Limits Sit Outside the Model > The wider technology and telecommunications week beyond the feature's lab findings: Domino Data Lab on fragmented internal AI access, Censuswide's finding that 76% of finance teams lack in-house agent expertise, AT&T's 450% agent-traffic figure, and BeyondTrust on machine identities outnumbering humans. Content type: article Source URL: https://www.agentpmt.com/articles/automation-and-technology-s-limits-sit-outside-the-model Markdown URL: https://www.agentpmt.com/articles/automation-and-technology-s-limits-sit-outside-the-model?format=agent-md Updated: 2026-07-26T12:28:15.619Z Author: Pancakes Tags: AI Agents In Business, AI Powered Infrastructure, Enterprise AI Implementation, AI Agent Identity, Security In AI Systems, News, Credential Vault --- Four separate datasets landed in technology and telecommunications this week, and not one of them was about model quality. They covered how employees reach AI inside their own companies, who in the building can explain what an agent did, how much network traffic agents generate per task, and what credentials they carry while they do it. Our feature this week goes deep on the lab evidence for what an agent does when its tools go quiet. Everything below is the wiring around that. * * * ## Every Business Unit Has Its Own Front Door to AI Domino Data Lab's fifth annual enterprise AI report, covered by Carrier Management on July 22, surveyed 639 senior AI leaders across North America, the UK, and continental Europe. Most coverage picked up the headline production and returns figures. The distribution finding underneath them got less attention and is arguably more actionable. 34% said their business units reach AI through inconsistent access methods, and 40% depend on a mediated layer to get there at all. Read that as an internal platform result rather than an AI result. A company where marketing hits a hosted model directly, support goes through a wrapper someone built last spring, and engineering runs its own gateway does not have one AI deployment. It has three, with three logging conventions, three credential arrangements, and three different answers to any question an auditor asks later. Mediated access is not automatically the weak option. A mediated path is frequently the well-governed one, since it puts a single choke point where policy, rate limits, and recording can live. Inconsistency across units is where the cost lands: no single record of what was asked, no comparable cost figure between teams, and no way to promote what one group learned to everyone else. For a technology or telecommunications operator running dozens of internal teams across network operations, billing, care, and field service, standardizing that path is ordinary platform engineering rather than an AI project. It pays twice. Consistent access raises adoption, because the route is documented and identical everywhere. It also makes an audit trail possible at all, because one path is one place to record what happened. Teams working the measurement side of this will find our longer treatment of [getting from pilots to a measured return](https://www.agentpmt.com/articles/the-agentic-roi-playbook-from-pilots-to-profit) covers the cost-per-outcome half. **Source:** Carrier Management * * * ## Three in Four Finance Teams Have No In-House Agent Expertise Censuswide's survey for Avalara, fielded June 15 to 22 across 1,505 CFOs and senior finance leaders at companies above $10 million in revenue in the US, UK, Australia, and India, carries a staffing number that deserves more attention than the governance headlines it generated. 76% said their organization lacks dedicated in-house expertise on how their AI agents actually function. Set that next to two other results from the same sample. 30% have not updated internal controls in the past year. And 92% reported career pressure to demonstrate a return on AI. Those three figures explain each other. A finance leader under pressure to show results, without anyone in the building who can walk through what an agent does step by step, has no practical route to writing a new control. You cannot write a control for a mechanism you cannot describe. So the controls stay as they were, and the agents ship anyway. The encouraging read is that this is a hiring and enablement gap rather than a technology one, and it prices a skill that is currently scarce. The person who can sit with a controller and reconstruct an agent run, which tools it called, what came back, what it wrote, and what it cost, is worth more to a finance organization right now than another model evaluation. That role does not require a research background. It requires knowing where the records live and being able to read them out loud. We made a version of this argument when [most of the Fortune 500 could not say what their agents did yesterday](https://www.agentpmt.com/articles/eighty-percent-of-fortune-500-companies-deploy-ai-agents-most-can-t-tell-you-what-they-did-yesterday), and the finance-side numbers turn it into a job description. One caveat on sourcing: Avalara commissioned the survey and sells tax compliance automation, so the governance framing serves its market. The methodology is disclosed and the sample is large, which is more than most vendor-sponsored research arrives with. **Source:** Avalara Newsroom * * * ## Agent Traffic Became a Capacity Planning Number On AT&T's second-quarter earnings call on July 22, reported in full by Benzinga, CEO John Stankey put figures on something wireless telecommunications companies have been sketching for two years. He cited industry research showing AI agents generate up to 450% more total traffic per task than a human performing the same work, and projected agentic adoption driving roughly 9x enterprise traffic growth and 7x consumer traffic growth by 2035. Attribute those carefully. They are Stankey's characterization of third-party research, delivered on an earnings call, rather than AT&T measurements, and the transcript carries no internal automation results to check them against. Read them as the planning assumptions a large carrier is willing to state in public. The mechanism behind 450% is familiar to anyone who has watched an agent run. A person looking something up issues one request and reads the answer. An agent issues a request, gets a partial result, retries, fans out to three more tools, re-reads its own context, and writes a record at the end. Every one of those steps is a round trip. Chattiness is a property of how agents work rather than a bug someone will patch out. Stankey's operational point stands on its own: networks now have to sense, decide, and act in near real time. That describes automation in telecommunications running inside the carrier at the same moment the carrier is hauling everyone else's agent traffic. For builders, this is a demand signal worth acting on. Traffic multiples of that size make request efficiency a design constraint with a price attached, and caching tool responses, batching calls, and picking tools that return complete results on the first try all move a number somebody is now measuring. A carrier planning capacity for an agent-heavy decade also means the headroom will be there, which is the more useful half of the forecast. **Source:** Benzinga * * * ## Machine Identities Now Outnumber the Humans Writing in The Hacker News on July 20, BeyondTrust's Len Noe made a point about population that reframes everything above it. Non-human identities, meaning service accounts, API keys, tokens, and certificates, now outnumber human identities in most environments. The directory that used to be a list of employees is mostly a list of software. Two references in the piece give that weight. Verizon's 2025 Data Breach Investigations Report puts stolen credentials as the most common initial access vector. And OWASP's 2025 Non-Human Identities Top 10 ranks improper offboarding first, ahead of every other failure in the category. That ordering is the useful part. Offboarding is an inventory exercise before it is anything else: knowing which credentials exist, which system each belongs to, when it was last used, and what happens to it the day the project that created it ends. Every organization already knows how to run that for people. Very few have run it for the software identities that now outnumber them. A ranked list from OWASP is the kind of clarity that makes a security topic tractable, and it arrived with a fix order rather than a warning. On the mechanism side, an agent never has to hold a credential at all. On AgentPMT, secrets sit in an encrypted vault and get injected server-side at the moment of use, with the scope attached to the connection instead of the agent. Offboarding then becomes revoking one connection rather than hunting for a key somebody pasted into a config two quarters ago. **Source:** The Hacker News * * * **Related:** Our feature this week, [Artificial Intelligence Technology Fabricates Missing Data](https://www.agentpmt.com/articles/artificial-intelligence-technology-fabricates-missing-data), goes deep on the lab evidence behind these findings. * * * ## Sources - New Research Finds That Despite Improved Productivity, AI ROI Fails to Outpace Spend, Carrier Management - Avalara Survey: Finance Leaders are Racing to Deploy AI Agents Before Governance is Ready, Avalara Newsroom - AT&T Q2 2026 Earnings Call: Complete Transcript, Benzinga - The New Insider Has No Pulse: Securing Privilege When the Actor Is an AI Agent, The Hacker News