Two numbers were published this year that look nothing alike. One was $355 million in internal AI-driven value at ServiceNow. The other was fifteen minutes saved per customer support ticket at BlackLine.
They are the same story told at different scales. In both cases the value came from a defined, repeated workflow that someone could count — not from a broad AI strategy, a pilot program, or a model upgrade.
That is the useful finding in the current evidence. The published wins are boring, specific and operational. The failures are usually ambitious, general and unmeasured.
What the public record actually shows
Five accounts from the past year give a reasonable cross-section. All are company- or partner-published, and that matters when reading them.
| Organization | Workflow | Stated result | Source type |
|---|---|---|---|
| ServiceNow | Internal IT and employee service | $355 million in AI-driven value; AI resolving 90% of some IT requests on first touch | Executive interview, company-stated |
| Smarsh (with Caylent, AWS) | Software delivery lifecycle | Up to $10 million in engineering and product capacity freed in year one; test coverage up about 30%; defect escape rate down 40% | Partner-published case account |
| BlackLine | Internal knowledge and content | 99% employee adoption; about 7,500 daily interactions; 15 minutes saved per support ticket; $200,000 annual savings on content and translation | Reseller-published case study |
| Zones (with Thoughtworks) | Enterprise order verification | Verification time down 50%, from up to a minute to 10–15 seconds; 17-month payback stated by the client | Consultancy-published case study |
| Pythian | Database operations and tickets | 80% faster incident resolution; threefold rise in active user engagement | Vendor blog, authored by the firm |
Read the third column carefully. These are self-reported outcomes with no independent audit, published because they worked. The organizations that tried the same thing and got nothing do not issue case studies. There is no denominator here, and no analyst has verified the arithmetic.
One illustration of why that caution is warranted: a Pythian customer account describes automating 10% of roughly 20,000 annual IT tickets, then attributes “1,000,000+ operational hours” saved to it. Those inputs do not reconcile — 2,000 tickets cannot produce a million hours at any plausible handle time. The rest of that account is specific and useful. That one figure is not repeated here.
Pattern one: value lands in a queue, not in a strategy
Look at where each organization pointed the technology. Zones had customer orders arriving with bespoke contractual rules that its order management system could not interpret, forcing customer success managers to hand-check up to 17 touchpoints per order. Pythian had 15,000 database tickets a month. BlackLine had roughly 70% of company data sitting in unstructured documents, recordings and transcripts.
None of those are exciting problems. All of them are countable. A queue has a length, a per-item cost and an owner, which means a change to it can be measured next month.
“Add AI to procurement” has none of those properties. It cannot be finished, so it cannot be evaluated.
Pattern two: the unit of value is usually seconds per touch
Almost every number in these accounts reduces to minutes or seconds at the level of a single transaction. Fifteen minutes per support ticket. Ten to fifteen seconds instead of up to a minute for an order review. Incident resolution time cut 80%.
That granularity is what makes the business case defensible. A finance leader can multiply a per-touch saving by a real volume and check it against the invoice. A claim that the organization is now “AI-enabled” cannot be checked against anything.
It also sets an honest ceiling. Seconds per touch is a good return on a high-volume process and a poor one on a process that runs forty times a year.
Pattern three: adoption is the delivery mechanism, not the afterthought
The BlackLine rollout is the most instructive account in this set, precisely because it is mostly about people. The company had already run an internal experiment with a self-hosted tool that demonstrated potential but could not scale. The Gemini Enterprise rollout that followed reached 99% of roughly 2,000 employees, with about 7,500 daily interactions making it the second most-used application after CRM.
The mechanism was deliberate: a company-wide introduction, then department-specific use cases, then executive workshops, then 40 small-group coaching sessions that were immediately oversubscribed. A prompt-a-thon drew 87 submissions from 131 employees.
The distribution of perceived benefit is worth noting. In a survey of its top 900 users, 45% said the tool saved them close to a day per week. That is a self-reported perception, not a time study — but it explains why the tool kept being used.
Contrast that with the standard pattern of buying licenses, announcing availability, and waiting for value to appear. Pythian describes that habit bluntly in its own account: chasing five-minute micro-efficiencies while missing structural workflow change. Its internal rollout produced a threefold rise in active engagement only after the supporting model changed.
Pattern four: deployment is the small half
Pythian's own framing is the most useful sentence in this body of evidence: deploying an agent is 20% of the journey, and maintaining accuracy in production is 80%.
The reason is mechanical. Models change, prompts drift, source systems get upgraded, and the documents a system was tuned against get replaced. An agent that measured well in month one can quietly degrade by month six while still producing fluent, confident output.
Zones built for that explicitly. The platform converts customer-authored standard operating procedures into machine-readable conditions, monitors incoming orders, and asks AI agents for recommendations. Crucially, the recommendations are staged in Azure SQL rather than applied. When a customer success manager opens an order, a panel shows the field, its current and proposed values, and the SOP evidence behind the suggestion. Approval writes the change through REST APIs.
That architecture is where most of the engineering effort went. Not into the model.
Pattern five: the confidence floor is a design decision
One line in the Zones account deserves to be copied into every AI business case: when confidence is low, no recommendation is generated, and the order follows the standard manual path.
That single rule converts an unbounded risk into a bounded one. An agent that must always answer will always answer, including when it should not. An agent permitted to abstain gives the business a measurable coverage rate — what share of cases it handles, and what share it correctly declines.
BlackLine's approach to its internal knowledge assistant reflects the same instinct from a different angle: answers grounded in designated source documents with inline citations, so a user can check the claim rather than trust the tone.
Pattern six: build a pattern, not a project
The accounts that describe durable value all produced something reusable. Zones describes a “procedure-to-action” architecture built to extend beyond order management without redesigning the underlying system. Pythian structured its work around 16 recurring agentic patterns and a repeatable operating loop, so the second use case costs less than the first.
The alternative — a bespoke build for each department — produces the same result twice at twice the price, and then three incompatible maintenance obligations.
Where the money is not
Four honest limits belong in the same conversation.
Low-volume work rarely pays. If a process runs a few dozen times a year, the per-touch saving is real and small, and the integration cost dominates.
Unmeasured programs cannot be defended. Nearly every credible account here began with a baseline: cost per order, tickets per month, minutes per review. Without one, the after-state is an opinion.
Payback claims need their terms. Zones states a 17-month payback; a Mexican retailer quoted through a Google Cloud keynote expects 10x return on a one-week build. Those may both be true, and they are not comparable — different denominators, different scope, different definition of return.
Attribution is the weakest link. When a support ticket resolves in a third of the time, some of that is the tool and some is the process redesign around it. Separating them requires deliberate measurement, not a dashboard screenshot.
How to find where AI pays in your business
The method that emerges from these cases is unglamorous and takes about a month to start.
1. Pick a queue, not a department. Find work that arrives, waits and gets handled: orders needing review, tickets awaiting triage, documents awaiting indexing, exceptions awaiting a decision. Write down its arrival rate and current handling time.
2. Count the touches per item. Zones needed up to 17 field updates before an order could be released. That number, multiplied by volume and handling time, is the size of the opportunity. It also tells you where the integration work will be.
3. Find the expert dependency. The strongest signal is work that only two or three experienced people can do correctly because the rules live in their heads or in documents nobody else reads. That is simultaneously the pain and the risk of losing them.
4. Decide the authority boundary before you build. Classify each action as read-only, reversible or irreversible. Recommendations can be generated freely; consequential writes need a named approver. Then set the confidence floor — the point below which the system abstains.
5. Instrument four numbers from day one. Time per touch, percentage handled without escalation, accuracy against a manually checked sample, and the share of cases the system declines. That last one is not a failure metric; it is how you know the boundary is working.
What to track in month six
- Hours identified versus hours realized. Zones identified 20,000 annual hours and expects roughly 14,000 to be realized under the current operating model. The gap is the honest number.
- Weekly active use, not licenses issued. BlackLine's 7,500 daily interactions matter more than its 99% adoption headline.
- Drift. Re-run the original test set quarterly. If accuracy fell, something changed and nobody noticed.
- Decline rate. A system that never abstains is not confident; it is uncalibrated.
The short answer to the question
AI is delivering today in back offices, service desks, order operations, document-heavy claims and compliance work, and in the supporting tooling around software delivery. It is delivering in seconds and minutes per transaction, at volumes high enough for those seconds to matter.
It is not delivering reliably as a general capability layered across an organization that has not decided what a completed task looks like.
The companies in these accounts did not succeed because they picked the right model. They succeeded because they picked a countable workflow, measured the before-state, built the operating layer nobody demos, and gave the system permission to say nothing when it was unsure. That is available to a business of almost any size, and it starts with a queue somebody is already complaining about.
Hashtags
#EnterpriseAI #CustomSoftware #SystemsIntegration #BusinessAutomation #CustomAI #AIAgents #WorkflowAutomation #OperationalExcellence #DigitalTransformation #AIGovernance #DataGovernance #ITLeadership #BusinessStrategy #RiskManagement #EnterpriseArchitecture #DataEngineering #CustomerExperience #ProcessAutomation #BusinessGrowth #MeasurableOutcomes