AI Workflow Automation

AT&T Runs 500 Agentic Workflows. The Count Isn't the Lesson

Rushil ShahRushil Shah
15 min read
Share

AT&T's near-500 production agentic workflows is being read as proof that agentic AI works at scale. It is better read as proof that a shared platform layer works at scale — and that is a different, harder, more copyable lesson.

TL;DR

AT&T's ~500 production agentic workflows is a platform metric, not an AI metric. You cannot reach three digits of workflows unless connectors, credentials, approvals, retries and logs are shared services rather than per-workflow code. Marginal build cost falls as the fleet grows; marginal maintenance cost rises. If under ~60% of your workflow components are shared, workflow #30 breaks you long before #500. And for most buyers the correct target is 20 governed workflows that own real state — not 500 thin ones.

Last week AT&T disclosed the largest publicly stated agentic workflow fleet to date. Nearly 500 agentic workflows are in production, the company processes about 45 billion AI tokens a day, its agentic platform handles more than 1 million transactions monthly, and the plan is to triple the workflow count next year. Within about 72 hours that figure was circulating as evidence that agentic AI has crossed into enterprise reliability.

It is evidence of something, but not that. Five hundred is a count of deployed artifacts. It tells you almost nothing about model capability and almost everything about the organisation underneath — specifically, that someone built a layer where a new workflow is an act of configuration rather than an act of engineering. That distinction is the whole article.

The Number That Matters Isn't 500

Do the arithmetic the coverage skipped. One million transactions per month across roughly 500 workflows is about 2,000 executions per workflow per month — roughly 66 runs per day for the median workflow. That is not a high-volume system per workflow. It is a very wide, fairly shallow one.

Now put the token number next to it. Thirty-three thousand platform transactions a day against 45 billion tokens a day implies about 1.35 million tokens per transaction, which is obviously wrong. The token figure is company-wide — assistants, coding, document and data tooling, the lot — and the transaction figure belongs to the agentic platform. The two numbers describe different systems and should never be multiplied together. I flag this because I have watched buyers build business cases on exactly this kind of cross-multiplication.

~500 agentic workflows AT&T says are in production, with plans to triple in 2027 Source: SiliconANGLE, 2026
1M+ transactions per month across the whole agent platform — about 66 runs/day per workflow Source: SiliconANGLE, 2026
mid-to-high 90s accuracy required before a workflow is allowed to run without a human in the loop Source: SiliconANGLE, 2026
40%+ of agentic AI projects Gartner expects to be cancelled by end-2027 on cost, unclear value or weak risk controls Source: Gartner via BigDATAwire, 2025

The honest reading of 66 runs per day is that AT&T has optimised for breadth of coverage over depth of autonomy, and that the economics work because the cost of adding workflow #451 is close to zero. That is the thing to copy. Not the number.

What AT&T Actually Built Underneath the Count

Three structural choices show up in AT&T's own description, and all three are platform decisions rather than model decisions.

First, few agents, central logic. The company deliberately breaks from the pattern of building large numbers of independent specialised agents, instead maintaining a relatively small set of agents that execute business workflows using centrally managed business logic held in context or memory, with retrieval agents finding applicable knowledge and planning agents deciding actions, tools and data. In practice this means a "workflow" is a configuration record — a process map, a policy, a tool allowlist — not a bespoke agent with its own prompt stack. That is why the count can run to 500 without 500 maintainers.

Second, a gated lane for anything that touches a customer. Employees can build personal productivity agents freely, but workflows affecting customers must pass through the Business Agent Foundry and clear security, privacy and compliance review. The foundry itself was stood up in early 2025 to work with individual business units and map processes against industry frameworks rather than collecting automation suggestions.

Third, autonomy as an earned state, not a build-time setting. Every production workflow starts with human oversight and only graduates once accuracy sits in the mid-to-high 90s within an acceptable risk profile, with continuous monitoring for drift. Earlier in the year, AT&T's chief data and AI officer described the same discipline from the platform side: a control agent coordinating smaller agents that typically run fine-tuned small language models because their tasks are tightly defined, with every agent action logged and data isolation, retention policy and role-based access enforced on every handoff.

Note what is absent from that list: any claim that a newer model unlocked the scale. The capability was adequate two model generations ago. The constraint was never the model.

The shared layer that makes workflow #451 cheap

1
Identity and credential broker

One place that holds every system credential, scoped per workflow. Rotating a mainframe service account should be one change, not 40.

↓
2
Tool / connector registry

Typed, versioned tool definitions — increasingly MCP servers — owned by the platform team, not copied into each workflow.

↓
3
Model gateway and routing

Single egress point. Model swaps, rate limits, fallback chains and per-workflow cost attribution live here.

↓
4
Approval and escalation service

Human-in-the-loop as a shared primitive with its own SLA, queue and audit trail — not a Slack message hard-coded into a node.

↓
5
Execution semantics

Idempotency keys, retry policy, dead-letter handling and compensating actions supplied by the runtime, identically for every workflow.

↓
6
Trace store and eval harness

Every tool call, argument and handoff recorded in one schema, replayable against a regression suite before any prompt or model change ships.

The Fleet-Scale Cost Curve Nobody Budgets For

Two curves move in opposite directions as a workflow fleet grows, and almost every budget I have seen models only the first.

Marginal build cost falls. Workflow #1 includes building the auth layer, the retry semantics and the logging schema. Workflow #40 includes none of those. On a mature platform the build cost of a new workflow converges on the cost of process discovery plus a config file — days, not weeks.

Marginal maintenance cost rises. Each workflow is a standing claim on a set of external interfaces. A fleet of N workflows depending on a pool of M systems accumulates breakage proportional to how often those M systems change, and the probability that at least one of them changed this week approaches certainty well before N reaches 100. The practical failure modes, in rough order of how often I see them:

  • Silent schema drift. A CRM adds a required field. Twelve workflows now fail validation at a step that used to pass. Nothing alerts, because the agent "handled" it by producing a plausible wrong answer.
  • Credential expiry cascades. One OAuth refresh token dies at 03:00 and twenty workflows go amber simultaneously. Without a credential broker, you find out from the business, one ticket at a time.
  • Prompt and model coupling. A model deprecation forces a swap; six workflows that were implicitly tuned to the old model's formatting behaviour regress. Without a replayable trace corpus you cannot tell which six.
  • Retry amplification. A downstream 500 triggers agent retries, which triggers more token spend and duplicate side effects. I have seen a non-idempotent "create ticket" tool open 40 tickets in nine minutes.
  • Approval queue collapse. Human-in-the-loop works at five workflows because one manager watches the channel. At fifty, approvals become the bottleneck and people start rubber-stamping, which quietly converts your governance control into theatre.

On the token side, the arithmetic is less frightening than people expect and the direction of travel is downward. Gemini 4 Hydrogen is priced flat across its 1M-token context at $1.50 per million input and $7.50 per million output tokens. Applying that input rate as a crude illustration, 45 billion tokens a day would be roughly $67,500 a day on inputs alone — on the order of $25M a year before output. That is my arithmetic on list prices, not AT&T's spend; a gateway routing most traffic to small fine-tuned or open-weight models lands far lower. The point is that inference is a line item you can forecast. Maintenance headcount is the one that surprises people.

Dimension5 workflows50 workflows500 workflows
OrchestrationWhatever the builder tool gives youShared runtime with versioning and environmentsRuntime + registry + policy engine as separate services
CredentialsStored per workflowCentral vault, per-workflow scopesBrokered, short-lived, auto-rotated, per-run attribution
ObservabilityExecution log, read on failureStructured traces + alerting on failure rateTrace store + replay + drift detection + per-workflow SLOs
EvaluationManual spot checksGolden-set regression per workflowShared eval harness gating every prompt/model change
OwnershipThe person who built itA small platform team + business ownersPlatform org, intake process, formal business case per workflow
Dominant costBuildBuild + reworkMaintenance, review capacity and change management
Failure you fearIt doesn't workIt breaks and nobody noticesIt works and does the wrong thing at volume
!

Treating the pilot count as the roadmap

Pilots get funded individually, so each one carries its own connectors, its own credentials, its own ad-hoc approval step and its own logging. That is rational for workflow #1 through #6 — the shared layer would cost more than it saves. The trap is that nothing in the pilot process ever triggers the decision to build it. Teams go from 6 to 18 workflows on momentum, and the first serious outage arrives as eighteen simultaneous incidents with no common trace format to diagnose them.

Fix: set the platform trigger in advance. At workflow #8, freeze new builds for one sprint and extract credentials, connectors, approvals, retries and logging into shared services. If you cannot afford that sprint, you cannot afford workflow #9.

The Reuse-Ratio Test

There is a quick diagnostic that predicts whether a fleet will scale, and it takes an afternoon. Inventory the components of your existing workflows — every connector, auth path, retry policy, approval step, logging call, schema definition, prompt fragment. Then count how many are shared versus per-workflow.

Below roughly 60% shared, the cost of workflow N grows with N instead of flattening, because every build re-solves problems someone already solved and every upstream change requires N edits. My experience is that teams in the 30-40% band hit the wall somewhere between workflow 25 and 35: the week where fixing existing workflows consumes more engineering time than building new ones. Above ~75%, adding workflows feels almost free and the binding constraint shifts to process discovery and review capacity — which is exactly where AT&T's constraint sits, given that its foundry works business unit by business unit and the company has had hundreds of business cases queued waiting for prioritisation.

Three sub-tests that sharpen it:

  • The rotation test. Pick your most-used system credential. How many files or UI panels must change to rotate it? One is healthy. Seven means you have no credential layer.
  • The deprecation test. Your primary model is sunset with 60 days' notice. Can you repoint traffic in one config change and run a regression suite to see what broke? If the answer involves opening workflows one by one, your model layer is per-workflow code.
  • The 3am test. A workflow fails at 3am. Is there a single place showing which step failed, which tool was called with which arguments, and what the model received? If observability is per-workflow, it is not observability.

What a Mid-Market Team Can and Cannot Copy

AT&T has a dedicated platform org, a formal intake office, and business cases tied to the CFO's office. A 200-person company has none of that, and copying the org chart is the wrong move. Copying the sequence is the right one.

Copy these. The gated lane — personal-productivity agents stay open, anything touching a customer or money goes through review. The autonomy gate — nothing runs unsupervised until measured accuracy clears a threshold you wrote down before you built it. The business case per workflow, however informal, because it kills the long tail of workflows that exist only because someone could build them. The centrally held process logic, so a policy change is one edit.

Do not copy these. The mainframe-integration pattern, unless you have a mainframe. The small-fine-tuned-model strategy: at AT&T's volumes, fine-tuning small models pays for itself; at 66 runs a day across 20 workflows it almost never does, and you inherit a retraining obligation forever. And do not copy the headline ambition — tripling a fleet is a sensible goal when your marginal build cost is near zero and a disastrous one when it is not.

The most transferable idea is actually the one AT&T's own lead stated as an endpoint: that in the long run the specialist foundry team should not exist and its work should be completely distributed. A central team is scaffolding for building the platform, not the permanent shape of the org. Smaller teams can skip straight to the end state if the shared layer is good enough — which is the premise behind how we structure automation builds and the integration work in our services.

The Honest Counterpoint: 20 Workflows That Own Real State

Here is the part that does not make headlines. For most buyers reading this, 500 is the wrong target and so is 100. The right target is roughly 20 workflows that own real state — that write to systems of record, make decisions with financial consequence, and have a named owner who would be paged if they broke.

A thin workflow summarises, drafts, routes or notifies. It is cheap, safe, popular with staff, and contributes a rounding error to margin. A thick workflow provisions the circuit, issues the credit, reschedules the engineer, adjusts the invoice. AT&T's reported wins are thick: changes to fiber installation orders that previously required pulling tickets from queues and manually updating records, and a halving of the time to provision dedicated internet in buildings with fiber already installed. Twenty of those beat five hundred summarisers, and the governance cost of twenty is tractable for a team of three.

Gartner's cancellation forecast is instructive here, because the stated causes were escalating costs, unclear business value and inadequate risk controls — all three are fleet-management failures, not capability failures. Projects die of operational debt, not of model limitations.

And there is a category where the answer is still "do not automate this yet": anything where the cost of a wrong action is irreversible and detection is slow. Agent capability has risen fast — one analyst briefing this month cited an autonomy horizon moving from seconds in 2020 to roughly 16 hours today, measured at a 50% success rate. Fifty per cent is a research milestone. Against a general ledger it is a restatement.

What This Week's Model Releases Change — and What They Don't

Two releases landed days before the AT&T disclosure and both are genuinely good for agents. Gemini 4 Hydrogen more than triples Terminal-Bench 4.1 scores over Gemini 3.8 Flash (57% vs 18.2%), cuts fabricated claims by more than 5x, and adapts when a tool call fails or a search returns empty — that last property is worth more to a workflow fleet than any benchmark number, because failed tool calls are the dominant runtime event at scale. Anthropic shipped Claude Fable 5.2 and Mythos 5.2 at a lower price than their 5.1 predecessors — $8 per million input and $40 per million output — generating up to 30% faster, with Fable positioned as an orchestrator that delegates simpler tasks to smaller models.

What they change: the floor on per-step reliability rises, and multi-model routing (big model plans, small model executes) becomes the default economic architecture rather than an optimisation. Google's own Orchestrator figures — near-top-model task success with 24% less quota and 19% faster completion by switching models within a session — are a preview of what every gateway will do by default.

What they do not change: none of these models rotate your credentials, none detect that your CRM added a required field, none run your approval queue, and none tell you which of your 40 workflows regressed when you swapped them in. Better models make each pilot easier to stand up, which means more teams will multiply pilot counts this quarter without a platform layer underneath — and will meet the wall at workflow 30 faster than the previous cohort did. If you are at 3-10 pilots and deciding what to commit to, the platform decision is the decision. Talk to us before you pick the tool.

Frequently Asked Questions

How many agentic workflows does AT&T actually run in production?

Nearly 500, according to AT&T Business AVP Luke Corbin speaking at an AWS event in Boston in November 2026, with plans to triple that in 2027. The same disclosure put the platform at over 1 million transactions per month and the company at about 45 billion AI tokens per day. Note that the token figure is company-wide across all AI use, not the agent platform alone — the two numbers should not be combined.

At what point do I need a platform layer rather than individual workflows?

Around workflow 8-10, in my experience. Before that, building shared credential, connector, approval, retry and logging services costs more than it saves. After about 10, every new workflow re-solves problems you have already solved and every upstream change requires N edits. The clean trigger is the first week your team spends more time fixing existing workflows than building new ones — if you wait for that signal, you are already a sprint late.

What is the reuse-ratio test for agentic workflows?

Inventory every component across your workflows — connectors, auth paths, retry policies, approval steps, logging calls, schemas, prompt fragments — and calculate what percentage is shared rather than per-workflow. Below about 60% shared, the cost of each new workflow grows with fleet size instead of flattening, and most teams stall between workflow 25 and 35. Above about 75%, adding workflows is cheap and your real constraint becomes process discovery and review capacity.

Why do agent pilots fail to scale into production fleets?

Rarely because of the model. Gartner's forecast that over 40% of agentic AI projects will be cancelled by end-2027 cites escalating costs, unclear business value and inadequate risk controls — all operational failures. In practice pilots die of schema drift that nobody detects, credential expiry cascades, retry amplification creating duplicate side effects, and approval queues that collapse into rubber-stamping once the volume of decisions exceeds what one human can genuinely review.

Should a mid-market company target hundreds of agentic workflows?

No. For most companies the right target is roughly 20 workflows that own real state — ones that write to systems of record and carry financial consequence — each with a named owner who would be paged on failure. Large counts are achievable only when marginal build cost is near zero, which requires a platform org most mid-market teams do not have. Twenty thick workflows beat five hundred thin summarisers on both margin and governance load.

When should an agentic workflow be allowed to run without human review?

Only after measured accuracy clears a threshold you defined before building, and only where the risk profile permits it. AT&T's stated rule is that every production workflow starts with human oversight and graduates to autonomy once accuracy reaches the mid-to-high 90s within an acceptable risk profile, with continuous drift monitoring afterwards. Where an error is irreversible and detection is slow — ledger entries, irreversible provisioning, regulated disclosures — keep the human regardless of the score.

Do the new agent-tuned models make fleet management easier?

They raise the per-step reliability floor, which matters: Gemini 4 Hydrogen reports more than 3x the Terminal-Bench 4.1 score of Gemini 3.8 Flash and better recovery when a tool call fails, and Claude Fable 5.2 shipped faster and cheaper than 5.1. But no model rotates credentials, detects a changed CRM schema, runs an approval queue, or identifies which of your workflows regressed after a model swap. Those remain platform problems.

agentic AIworkflow automationAI platform engineeringenterprise AIAI governanceobservabilityAT&Tn8n

Published

AI-assisted writing · Reviewed by the Twarx research team

Share:
Share

Research digest

AI Research Briefing

Honest insights on AI agents, Small Language Models, and local RAG. No hype. Only when we have something worth sending.

  • No hype, just measurable outcomes
  • Read by 2,400+ engineers
  • Unsubscribe anytime

Continue reading

More from Articles

AI Workflow Automation

Where Finance Agent Handoffs Break: A 2026 Six-Layer Fix

17 min read
AI Workflow Automation

Procurement's Six Seams: Intake, Vendor, Budget, Approval

14 min read
AI Workflow Automation

Procurement Agents: Requisition to ERP Without Dropped State

16 min read
AI Workflow Automation

Confidence Gates: How 14,000 Invoices a Month Get Reviewed

16 min read
AI Workflow Automation

AP and Financial Close: Designing the Agent Handoffs

16 min read