Let the numbers land one at a time. Stop on 37%: everyone uses AI, few see value. Then the line that frames the whole briefing: the constraint has moved inside the enterprise.
The executive map
8 changes. 6 decisions.
Six call for a leadership decision this quarter. Two are watch areas. Each one feeds your 90-day agenda.
Model strategy is becoming a routing and economics discipline. A single standard model tends to overpay for routine work and underperform on complex work, and at enterprise volumes that gap can be a large share of the AI budget.
Decision 1 of 6 · how to start
Set up task-level model routing
For your three highest-volume AI workflows, define the quality each one requires. Then test at least two providers, three effort settings and one open-weight model, using your own data. Review the routing every month.
Suggested owner: CTO
The research
Two trends are running at the same time. Epoch AI estimates that reaching a given level of AI performance gets about 13 times cheaper each year. A separate academic study finds that running the very newest models keeps getting more expensive. Routing is how a company captures the first trend without paying for the second.
A single model's cost can vary by up to 15 times depending on how much reasoning effort it is allowed to use. Often the cheapest option that meets the quality requirement is a frontier model at a low setting, or a smaller model at a high one.
Open-weight models now score about 12 points below the top of the benchmark at a small fraction of the price. MiMo-V2.6-Pro, for example, scores 46 at $0.13 per task, compared with 58 at $5.98 for the top setting. Because these models can run on your own infrastructure, they also address most data-residency concerns and make costs more predictable at high volume. On Vercel's gateway in June, they handled 29% of traffic for less than 4% of spending.
Diagram facts for the illustrative invoice-dispute workflow.
Prompt
Loop
Graph
What it does
Improves one interaction
Automates a job
Redesigns a workflow
Model calls
1
1 + retries
1 + retries per loop
Checks
0
1, with a stop condition
2 loops
Systems and records
0
0
3
People
0
0
1
Permission gates
0
0
1
The explanation
Most of the reliability gain comes from one step: the system checks its own work and tries again when the check fails. In one study, that step alone raised accuracy from 19% to 44% using the same model.
3 in 4of the companies getting the most value from AI have redesigned their workflows, compared with 1 in 4 of other companiesSurveyMcKinsey & Company · Aug 2026
Much of the reliability gain comes from the loop, not from the workflow around it. If loops are connected into a larger workflow before each one works reliably, their errors compound.
Decision 2 of 6 · how to start
Redesign one cross-functional workflow
Choose one workflow that spans at least two teams. Make sure each automated step checks its own output against a written definition of a correct result, and only then connect the steps into a single workflow.
Suggested owner: COO
The research
A prompt produces an answer. A loop checks that answer against a reliable reference, such as the purchase order, the contract or a test, and tries again until it passes or reaches a set limit. A workflow, sometimes called a graph, connects several loops with systems of record, approvals and people.
The evidence supports building in that order. In one coding study, adding a test-and-retry loop raised accuracy from 19% to 44% with the same model. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027 because of cost, unclear value or weak risk controls.
Nearly three quarters of the organizations getting the most value from AI have redesigned their workflows. Among all other organizations, about a quarter have.
Read the three statements as you move through the states. In the loop state, point to the check and the stop condition. That is where the reliability comes from.
03Change 03 · DecisionsRoutine decisions can run on AI
Automation creates capacity. What leaders do with it creates value.
The fact
$0.02
to process 1,000 routine decisions with a purpose-built decision model
63→95%accuracy on one task when a single question was broken into five smaller onesResearchBeri · Sep 2026
What this means
First-order return
Do the same work with less.
Second-order return
Do more with the same.
Strategic return
Do something you couldn't do before.
Decision 3 of 6 · how to start
Decide where AI-created capacity goes
Choose one high-volume process. Automate the routine work. Measure the capacity released. Decide where it goes.
Reduce cost
Absorb growth
Elevate human work
Create new value
Suggested owner: COO
The research
At list prices, 1,000 short decisions (about 500 tokens of input each) cost $0.02 with Jev, $0.06 with GPT-6 Luna, $0.60 with Claude Haiku 4.5 and $2.40 with Claude Opus 5.5. These prices say nothing about accuracy, which has to be measured on your own cases.
Accuracy depends largely on how the decision is set up. In an independent test, Jev was 62.6% accurate when asked a single question and 95.0% accurate when the same decision was broken into five questions weighted using labeled data. Test on 1,000 to 2,000 past decisions and set a separate threshold for each question; where to set it is a business decision about acceptable error, to be documented and reviewed like a credit limit.
IKEA's largest retailer, Ingka, reports that Billie handled 47% of customer queries to its call centres over two years. Over the same period it trained 8,500 call-centre workers as remote interior design advisers, a channel that brought in €1.3 billion in financial year 2022 (3.3% of revenue), with a goal of 10% by 2028. Verizon gave its 28,000 service reps a Gemini-based assistant from July 2024; its consumer CEO says sales through the service team rose nearly 40% after deployment. Neither company attributes these results to AI alone.
Start at ‘Balanced’: of an 83-person invoice-review team, about 73 people's time comes back, for an AI bill of about $50 a year (about $6K even on a frontier model). Then the turn: time saved is not value created. Reduce and absorb are legitimate; elevate and create are where IKEA and Verizon took it. Close on the decision: measure the capacity, then decide where it goes.
04Change 04 · VoiceThe interface recedes
Voice is becoming a practical way to get work done.
Simulated request: "Move tomorrow's client meeting to the afternoon and send them the updated prep package." Timings are representative, not measured.
Step
Stage
Start (ms)
End (ms)
What happens
Listen
Transcription
0
3500
Audio streams in while the person speaks.
Transcribe (streaming)
Transcription
200
3600
Speech becomes text as it arrives, so the system can start before the sentence ends.
Understand intent
Conversation
3600
3780
The sentence is turned into two specific actions, each with the details needed to carry it out.
Find the right records
Conversation
3780
3990
The calendar, the CRM and the document library are searched, so ‘the client’ and ‘the prep package’ are matched to real records.
Check permissions
Transaction
3990
4060
The assistant is allowed to reschedule meetings the company organizes and to share documents already approved for this client. Anything outside those rules would stop here and go to a person.
Update calendar
Transaction
4060
4330
The meeting moves to the first afternoon slot that works for every attendee.
Retrieve documents
Transaction
4330
4470
The latest approved version of the prep package is selected, not an old draft.
Prepare communication
Transaction
4470
4760
A short note to the client is drafted from an approved template, with the new time and the package attached.
Complete and confirm
Transaction
4760
5200
The message is sent, every step is logged, and the person hears a summary of what was done.
The explanation
Speech models now answer in about a second. The value is one spoken request that completes a task across company systems.
125,000+businesses use Wispr Flow for voice dictation, including most of the Fortune 500PressPulse 2.0 · Aug 2026
Speaking takes 3.6 of the 5.2 seconds. The value is in the 1.6 seconds after, when the system finds the records, checks permissions and makes the updates. Voice pays off mainly where those connections already exist.
Decision 4 of 6 · how to start
Choose one transaction to run by voice
Pick one transaction where typing or navigating screens slows people down, in field service, sales, operations or customer service. Build it end to end, including the permission check and a handoff to a person.
Suggested owner: Chief Customer Officer
How people use voice today
Personal productivityMost business value →
DictationEmails, documents and AI prompts, spoken. About three times faster than typing.
The foot pedalHold to talk, release to send, with hands kept on the keyboard.
Thinking out loudVoice mode in ChatGPT, Claude or Gemini to prepare and draft on the move.
Hands-busy workTechnicians, clinicians and drivers update records without a screen.
Customer voice agentsService lines that resolve simple requests, with a person on hand.
Requests that get doneOne instruction that updates several systems. Shown below.
The leading speech models now start responding within 1.1 to 1.4 seconds and score above 90% on a spoken-reasoning benchmark. That is close to the pace of normal conversation, which makes voice practical for customers, field technicians and sales teams.
Most of the work happens after the person stops talking. The request has to be turned into a structured instruction, matched to the right records, checked against policy, recorded in the system of record and confirmed back to the user. The trace above shows each step and how long it takes.
Customer service leaders increased AI spending by 38% while their overall budgets grew 2%. Customers have one condition: 87% want to be able to reach a person.
AWALI framework. Cost bars in the figure are relative, not measured.
Workflow type
Examples
Strategy
Why
Systems of record
General ledger, payroll, ERP core
Buy
Every company needs the same thing here, and the vendor carries the compliance burden. Buy it and manage it through procurement.
Standard workflows
HR service desk, expense approval
Configure
The process is much the same across companies. Buy a platform, configure it to fit, and make sure your data stays portable.
Competitive workflows
Pricing approval, field dispatch, underwriting
Bespoke
How you run this process is part of how you compete, and off-the-shelf software forces you to work like everyone else. With AI, building it yourself now makes financial sense.
Core judgment
Your models of risk, demand and customer value
Own
This is where your advantage comes from. Own the whole system, keep it out of vendors' training data, and invest most heavily in testing.
The explanation
AI has cut the cost of writing software, so building is viable again for the workflows that set you apart. It has not reduced the cost of testing it, maintaining it and owning it over time.
−19%measured change in speed for experienced developers using AI, even though they believed they were 20% fasterExperimentMETR · Jul 2025
What this means
The real difference is what the software touches. Tools that sit on top of existing data are low risk to build in-house. Anything that writes to systems of record, touches customers or money, or has to behave the same way across the business needs a proper data and governance foundation first.
Decision 5 of 6 · how to start
Re-evaluate one differentiated workflow for build versus buy
Pick one strategically differentiating workflow that a SaaS product constrains, and compare building and buying on full cost, including testing, maintenance and ownership. Decide which parts your team builds and which foundation layers, such as data and governance, come from a specialist partner or platform. Apply the same review to software renewals above a set value.
Suggested owner: CIO
The research
The research on developer productivity is mixed, and the differences are informative. On new, well-defined tasks, developers were 21% to 55% faster. In METR's 2025 trial, experienced developers working in their own mature codebases were 19% slower, even though they believed they were 20% faster. Google's DORA research links AI use to faster delivery but less stable systems.
The practical conclusion is to build where the logic is specific to your business and the scope is clear, and to fund the review, testing and ownership needed to keep it reliable.
Figure 05b
The productivity gain depends on the task, and developers tend to overestimate it
Reported effect of AI coding tools in controlled trials. Positive values mean faster work or more output. Each study measures something slightly different.
Move from ‘Buy’ to ‘Own’. Then turn on ‘Show the build cost before AI’: the dashed outline shows why building wasn't viable a year ago. Notice that the testing segment didn't shrink.
06Change 06 · CreativeCreative becomes a production system
Creative AI is becoming a production system.
The fact
~$3–12
base model cost to generate 30 seconds of video, before editing and review
Write a short shot list and set the look with one key frame from an image model.
Use saved camera presets instead of describing the move in words.
Send wide shots and B-roll to a lower-cost model; keep the best model for the one or two hero shots.
Train a Soul ID if the same presenter or character appears across the campaign.
Generate versions by format and message, then cut, grade and caption in your usual editor.
Test the versions and track the cost of each ad that beats the current one.
Clips run up to 15 seconds per generation, final assembly still happens in an editor, and the capability descriptions are Higgsfield's own.
What it does well
Many models, one placeVeo 3.1, Kling 3.0, Seedance 2.0, WAN 2.6 and others sit alongside Higgsfield's own models, so each shot can use the model that suits it.
Camera controlDolly, orbit, tracking and crane moves are set when the clip is generated, rather than hoped for in a text prompt.
Consistent peopleSoul ID is trained once from 20 or more photos and keeps the same face across every clip and model.
From product page to adMarketing Studio reads a product image or link and fills templates for product shots, Meta and Google ads, short UGC-style video and motion graphics.
Editing in placeInpainting, relighting, upscaling and layers change a scene without regenerating it from scratch.
Platforms such as Higgsfield put the leading video and image models in one workflow, so a small team can test far more creative. Higgsfield adds camera control, consistent characters and ad templates on top of the models. A small team can now produce and test more versions than a traditional production schedule allows.
1,500+ad templates in Higgsfield's Marketing Studio, which builds ads from a product image or product-page linkVendor-reportedHiggsfield · Aug 2026
What this means
When producing a campaign's worth of versions takes days rather than months, the useful measures change: the cost of each ad that beats the current one, and how quickly results feed the next brief.
Watch area · how to start
Baseline creative testing speed
Measure how many ad versions you test each month and how many days it takes to act on results. Then run one bounded test in a single channel with a production platform such as Higgsfield. No enterprise commitment is needed yet.
Suggested owner: CMO
The research
Generated video lists at $0.10 to $0.40 a second, so 30 seconds costs $3 to $12 in model fees before retries, editing and review. Production cost is no longer the main constraint.
Audience reaction still has to be tested rather than assumed. In IAB's survey, 82% of advertising executives believed young consumers feel positive about AI ads; 45% of those consumers do.
Material AI changes in 2026 to 2 October. Kling 3.0 is dated to February 2026 by Kuaishou; the day shown is approximate. Anthropic's >80% figure is as of May 2026, published June 2026. Model releases are the flagship releases of OpenAI, Anthropic, Google, xAI and Meta plus the leading open-weight releases; dates follow public release trackers, using the first public availability where trackers differ.
Under an annual plan, each waited about 204 days for a decision. With a monthly review, the wait drops to 13. AWALI tracked them through September 2026, including 25 major model releases: more change than a once-a-year plan is designed to absorb.
~4.3monthsfor the length of task AI can complete to double, based on independent measurementBenchmarkMETR · Jan 2026
What this means
Annual planning still works for setting strategy, and quarterly cycles still work for allocating capital. AI operations increasingly need a decision point every month.
Watch area · how to start
Instrument your AI planning cadence
Log material AI developments and how long each waits for a decision. Set a trigger: if more than one a month affects your priorities, add a 30-day AI operating review alongside annual strategy and quarterly capital allocation.
Suggested owner: COO
The research
The length of software task AI can complete went from about an hour in early 2025 to about 12 hours by February 2026. Since 2023 it has doubled roughly every 4.3 months. METR notes that its tests can barely measure the leading models, and that the task length at 80% reliability is much shorter.
AI labs report that AI now writes most of their own code: more than 80% of merged code at Anthropic and 75% of new code at Google. Both figures are self-reported and measured differently.
Something that was out of reach when the budget was set can become routine before the end of the year. A monthly review needs a clear trigger for revisiting work when that happens.
Figure 07b
The length of task AI can complete doubles every few months
METR's measure of the length of software task AI can complete with a 50% success rate, by model release date, on a log scale with 95% confidence intervals.
Illustrative scenario: one operations agent with access to Email, CRM, Finance, Documents, Calendar, External tools, Customer comms. Nothing is executed.
Control
When on
When removed
Identity
Every action is signed by the agent's own identity, with a named owner.
Actions run on a shared login, so nobody can tell which agent did what.
Permission scope
The agent can reach only the systems and actions this job requires.
The agent has broad access, so deleting records, wiring money and calling any API all become possible.
Approval thresholds
External emails, file shares and customer messages need a person's approval.
Actions that can't be undone go ahead without anyone approving them.
Spending limits
Payments are capped at $1,000, and the agent can't make purchases or issue refunds.
There is no limit on how much money the agent can move.
Rate limits
The agent can take at most 20 actions a minute, so mistakes stay small.
The agent can take thousands of actions a minute, so one bad instruction can repeat many times before anyone notices.
Audit logging
Every action is recorded with who did it, what went in and what happened.
Nothing is recorded, so incidents can't be reconstructed.
Human escalation
Unclear or out-of-policy cases go to a named person.
The agent decides for itself how to handle exceptions.
Kill switch
An operator can stop the agent and revoke its access in the middle of a task.
Once a task is running, it can't be stopped.
The explanation
Enterprise exposure is determined less by an agent's raw capability than by the authority, access and controls surrounding it. Agents can now send email, move money and change records on their own. A written policy is unlikely to stop one in the middle of a task; controls built into the systems it uses can.
>40%of agentic AI projects are expected to be canceled by 2027, partly because of weak risk controlsForecastGartner · Jun 2025
What this means
Only 34% of organizations apply the same controls to AI agents as to employees. The gap between a typical pilot and a governed agent is four controls, and each is far easier to put in place before launch than after an incident.
Decision 6 of 6 · how to start
Set real-time controls for AI agents
Make an inventory of every AI agent and automated account. Give each one a named owner, its own identity, spending and activity limits, approval thresholds, logging, an escalation path and a tested way to shut it down.
Suggested owner: CISO
The research
The necessary controls are well understood: a separate identity for each agent, only the access it needs, approval thresholds, logging, ongoing testing and the ability to revoke access immediately. Recent incidents show why they need to be enforced while systems run rather than written into policy. In 2025 a coding agent deleted a live production database during a code freeze, and AI carried out 80% to 90% of the work in a state-linked hacking campaign. In 2026, evaluation agents reached a third party's production systems.
Practice is behind. Only 34% of organizations apply the same controls to AI agents as to employees. OWASP's first list of top risks for agentic applications includes identity abuse, misuse of tools and agents acting outside their instructions, and NIST has started standards work on agent identity and authorization.
Start with ‘Fully governed’. Switch to ‘Typical pilot’: identity, limited access and a log are on, yet the agent can message customers and issue refunds unsupervised and can't be stopped. Then ‘No runtime controls’: deleting records and wiring money open up too.
Your 90-day operating agenda
You now have the eight changes. Here are the decisions to make.
Each of the six decisions has a suggested owner. In an AWALI workshop, we turn them into named owners, dates and a first operating review within 30 days.
One author · 65 sources · 44 AI developments · 8 changes · 6 decisions
Conceived, researched, written and built by Scott ThomasAWALI · Current to 2 October 2026
Next briefing
The next one, the day it is published.
One email per quarter, when a briefing is published. Nothing else.
You are on the list.
The next briefing will reach you the day it is published. Until then, the full library is at awali.tech/thought-leadership.
Finish on the agenda. Assign owners before the meeting ends, and send the brief the same day.
Sources and method · 65 sources
Where every figure comes from
Figures are reported as published by each source as of the date shown. This edition was compiled on 2 October 2026, and model and pricing data reflect that date.
Where a shorter summary would overstate a finding, we use the source's original, narrower wording. AWALI has not independently reproduced these studies.
Self-reported figures, press reports, forecasts and AWALI calculations are labeled wherever they appear.
Every chart is drawn from the values shown in its table view. Figures we could not confirm were left out rather than estimated.
The interactive scenarios (the routing policy, the decision simulator, the voice trace, the build-versus-buy comparison, the ad-testing example and the agent controls map) illustrate how these systems behave. They are not measurements.