Co-Founder & CEO of CLRT · Dubai
Vishal Sachar
Vishal Sachar is the Co-Founder and CEO of CLRT, where he helps UAE businesses make sense of applied agentic AI and put it to work. He writes on agentic systems, AI governance, and the economics of automation.
At CLRT, Vishal leads the firm’s strategy, product direction and client relationships. His work starts with how a business actually operates: where an agent can create value, where judgement must remain human, and what it takes to turn an experiment into dependable work.
Writing by Vishal Sachar
Field notes on the decisions behind applied AI, drawn together here from the CLRT Insights archive.
· Fundamentals
Two Thousand Fake Cases
The profession filed hallucinated citations under 2023, an embarrassment the professional tools were supposed to have fixed. The public database of court decisions stood at 2,041 entries on 14 September 2026, 2026 has already outrun the whole of 2025, and the sanctions have moved from fines to a struck-off lawyer.
· Method
The Prompt Course Is the Wrong Purchase
Companies buy prompt-engineering courses as if capability came by the seat. The measured evidence says the syllabus barely moves the output, and that what pays is the part no course contains: the organisation's own requirements, its workflows, and the rule for what gets checked.
· Method
The Meter Is Also a Model
Outcome pricing promised to move the risk onto the vendor. What actually moved was the audit. The vendor now defines the outcome, reads it with its own model, and three vendors have rewritten their definition this year.
· The Market
The Price Cut Is the Warning
August's price cuts read as good news for buyers. Ramp's spend data says why the labs are cutting: businesses declined the frontier premium and the heaviest spenders trimmed. A price war is how procurement discipline arrives, not how value does.
· Sectors
Thirty Minutes to Five
Dubai's property regulator has automated the step every developer used to wait on. That makes the private firm's own back office the slowest node in a transaction, and the launch figure everyone is quoting is the wrong number to worry about.
· Fundamentals
Your Model Agrees With You. That Is the Bug.
Executives read a model's agreement as validation. The measured evidence says agreement is a trained artefact, and it rises with exactly the habits senior people bring to the conversation: pushing back, pulling rank, giving orders, showing strain, and letting the assistant remember them.
· The Market
Thirty-Two Advisors in the Cabinet Room
On the day Dubai's private sector was handed a training course, the UAE Cabinet put thirty-two AI advisors inside its own decision process and published the numbers it will judge everyone else against. The gap that opened is not capability. It is measurement.
· Fundamentals
Half the Time Is Not the Same as Every Time
The market reads a twelve-hour time horizon as an agent doing a day's work. The lab that publishes the number says it measures a coin flip, and the lab's own data file puts the reliable figure for the same model at about seventy minutes.
· Method
The Divergence Is the Test
CLRT closed its AI After Hours series with a hackathon built around a company that does not exist, seeded with six planted problems. The room was scored on which problem it chose, how it defended that choice to a board, and what it admitted its build could not do. Build speed was never a criterion.
· Fundamentals
The Window Got Bigger. The Memory Did Not.
Context windows grew from thousands of tokens to millions, and the market read the number as a cure for the assistant that loses the thread. Every benchmark that measures it finds the usable window is a small fraction of the advertised one, and memory bought the same way fails the same way.
· Focus
Twenty-Three Asks, Five Machines
We asked a beginners' room in Dubai what it wanted to learn, and it said agents. Then we asked for the one task each person would never do again, and got twenty-three answers that needed almost none. The gap between the two lists is the whole job.
· Sectors
The Exception Was the Whole Year
The ports sector tells its AI story in berth optimisers and thousands of digital workers. The 2026 half-year filings say the result was made by people moving a nation's cargo around a closed strait, and neither operator has published a single AI outcome figure.
· Fundamentals
The Reviewer Fails on Schedule
Every AI governance policy ends with a person who signs off the output, and nobody asks how reliable that person is. The research has been asking for thirty years. The answer is a curve, not a safeguard, and it bends the wrong way.
· Sectors
The Central Bank Already Wrote Your Agent Policy
Banks and insurers in the UAE talk about agents as something they are waiting to be allowed to run. The regulator settled that question in February. What it asked for in return is a document most licensed institutions cannot currently produce.
· Sectors
Decision Support, Not Decision Making
Abu Dhabi holds more connected clinical data than almost any health system on earth, and its regulator has said exactly where AI may stand in relation to a surgeon. Read the two together and the safe, measurable work for a hospital operator is not where the announcements point.
· Focus
What 300 Registrations Says About Dubai
We put a beginners' building night on Luma and watched three hundred registrations arrive in under a week, the first two hundred inside three days. The number is not a marketing win. It is a diagnostic of where Dubai actually sits with AI, and it contradicts the story most organisations are telling themselves.
· Building
The Data Debt Comes Due
Your data was always worse than the dashboards admitted, and nothing broke, because every person who used it carried a private map of its lies. The agent arrives without the map.
· Fundamentals
Small Models, Big Jobs
Agent economics do not improve by waiting for the frontier to get cheaper. They improve when someone decides which model calls never needed the frontier at all, and makes that decision hold as engineering rather than instinct.
· The Market
Sovereign Compute Is Not Sovereign Capability
The UAE has bought the most visible layer of the AI stack at a scale few nations can match. The two layers that turn gigawatts into advantage cannot be bought in any deal.
· Fundamentals
The Token Price Paradox
The price of a token has fallen faster than compute or bandwidth ever did, and AI bills are rising anyway. The paradox comes apart the moment you stop watching the price and start watching what the tokens are being spent on.
· Method
Your Company Already Adopted AI. Nobody Told You.
The adoption debate your leadership is still having was settled months ago by the workforce, one personal login at a time. What remains undecided is whether the company will ever see the adoption it already has.
· Sectors
What Klarna Teaches, and What It Doesn't
Customer support is the function everyone automates first, and Klarna ran the experiment in public. The market took two opposite lessons from the result. Both are wrong, and the useful one is quietly transferable to every function you run.
· Building
The Browser Is the Worst Room in the House
The agentic browser sends the most trusting employee you have ever hired into the most adversarial environment ever built. What gets disclosed as bugs is really the physics of the medium.
· The Market
Where the Money Actually Lands
The keynote says AI is transforming every function. The budgets say something else. Follow the composition of enterprise AI spend and a different map appears, one where the money pools in the functions that already knew how to buy software, and the functions with the most to gain carry single digits.
· Method
The Agent Is a New Joiner, Not a New Tool
Nobody gives a new hire the master keys on day one. An agent gets them on install. The mental model that fixes this is one every company already runs.
· Fundamentals
From Seats to Outcomes
When software prices per outcome, it is priced like labour, and procurement has no muscle for that. Seat pricing made cost predictable and value vague; the outcome reverses both, and the reversal is the buyer's new job.
· The Market
The Pilot Graveyard
The number everyone quotes says AI pilots fail. Read the study behind it and it says something more uncomfortable: most companies choose pilots that were never pointed at anything worth moving.
· The Market
Ninety Percent Expected It. Eleven Percent Shipped It.
Four independent surveys wrote the 2026 ledger on AI agents: expectation booked at ninety, delivery shipped at eleven. The gap is not the one the market thinks it is, and it is good news for buyers.
· The Market
The Model You Rent Can Be Recalled
A frontier model launched on 9 June, was suspended worldwide by a US export-control order on 12 June, and returned on 1 July. If your operations are hard-wired to one vendor's model, that is a class of continuity risk no uptime SLA covers.
· Method
The Token Bill Is a Judgment Bill
Uber put coding agents in front of roughly 5,000 engineers and exhausted its entire 2026 AI budget by April, in the very window tokens became cheaper than ever. The budget did not break because models got expensive; it broke because nobody priced the work.
· The Market
The Deadline Arrived Before the Rulebook
The UAE has put dated deadlines on agentic AI adoption while the rules that will govern it remain unwritten. Most firms read that gap as a reason to wait. It is the opposite: it is the brief.
· Building
The Uninsured Agent
Since January 2026, insurers have been able to write generative AI out of standard liability cover, and the agent incidents they are backing away from have stopped being hypothetical. Most businesses running agents today are carrying that risk on their own balance sheet, and nobody has told them.
· Focus
Good, Not Leverage
Ascent finds the work you are both good at and energised by, then weighs it against the goal you actually stated. When a beloved task does not move that goal, it says so, in three words no readiness tool is built to say.
· Focus
The Verification Tax
Delegation to AI is deepening on hard telemetry, not survey sentiment. But the manager's week does not shrink, it converts: hours saved doing become hours spent verifying, and almost nobody is budgeting for them.
· Focus
Your Team Feels Faster. The Stopwatch Disagrees.
The only randomised trial of experienced developers using AI on their own repositories measured a 19% slowdown. The developers came out believing they had been 20% faster. Most companies allocate budgets on the second number.
· Focus
The Unallocated Hour
AI verifiably hands time back, and the best-instrumented study of AI at work finds the effect on earnings and hours is roughly zero. The reason sits in one number: 85% of users pour the freed hour straight back into the job it came from.
· Building
Loop Engineering
The most advanced AI practitioners have stopped being the person who prompts the agent. They design systems that prompt for them, moving the leverage point from the prompt to the loop.
· Building
The Plan That Survives February
A plan does not die in February from weak discipline. It dies because of where it was stored: your own memory, the one thing you cannot scale. Built like a good agent, with a memory outside your head and a loop that surfaces the next action, it remembers itself.
· Focus
The Week Is the Diagnostic
Every serious diagnostic method commits to a unit of analysis. For agentic AI, the right one is deliberately small: one person, one workflow, one representative week, because the week is where the evidence actually lives.
· Focus
Decision Fatigue Is Dead. The Fourth Hour Is Not.
The willpower science that shaped a generation of executive calendars collapsed under replication, and almost nobody running a company has noticed. What survives is narrower, better evidenced, and far more useful in a company run on agents.
· The Market
The Constraint That Is Also a Moat
Where your data is processed is a decision most businesses make by accident. Make it on purpose and the residency rule that constrains you becomes a wall your competitors cannot climb.
· Focus
The Twenty-Three Minute Myth
The most quoted focus statistic in the world, that it takes 23 minutes to recover from an interruption, has no paper behind it. What the research actually found is stranger, and the real 2025 numbers are worse than the myth.
· The Market
What the Dubai AI Seal Actually Measures
Since a late-2025 directive steered government buyers toward certified AI suppliers, the Dubai AI Seal stopped being a badge and became a gate. Here is what it actually rewards.
· Building
Building a Company of Agents
You do not manage a company of agents, you specify it and you verify it. That makes the skill closer to engineering intent than leading a team, and it leaves one true ceiling on how large you can grow: how much you can check.
· Method
CLRT Ascent
Ascent turns where AI belongs in your work into a number and a plan, quantified in dirhams. Its defining feature: it will tell you a task you love is not your leverage.
· The Market
Redeployment, Not Reduction
In most markets the AI business case is sold as headcount reduction. In the UAE that framing is not just distasteful, it is strategically wrong and it will cost you the room.
· Method
Governance Is What Lets You Look Away
Governance sounds like the committee that slows things down. It is the opposite. It is the only thing that lets you take your hands off an AI agent at all.
· Method
Build, Adopt, or Walk Away
Every AI opportunity sorts into one of four boxes: build now, fix the organisation first, resist a tempting waste, or walk away. The most valuable box is the one that says no.
· The Market
From Federal to Private
The pressure behind Dubai's private-sector mandate is not coming from the private sector. It is coming from a government that has already started moving and is not waiting for anyone to catch up.
· The Market
The Capability Gap
Almost every organisation now uses AI somewhere. Almost none capture real profit from it. The distance between those two facts is the entire opportunity, and the entire risk.
· The Market
The Agentic Mandate
Dubai is the first economy on earth where adopting agentic AI is a government directive with a deadline, not a market trend. That single fact reorders every priority for a business here.
· Building
Engine-Agnostic by Design
A new "best" model arrives every few weeks. Hard-wire yourself to one and you choose between endless rebuilds and falling behind. There is a better posture, made once.
· Building
Evals Are the New Tests
Traditional tests assert exact outputs, but AI agents are probabilistic. The new core skill is writing evals that ask whether the result meets the goal.
· Building
Agent Skills
If you explain your business to an AI from scratch every time, you are paying a tax you do not need to. The fix has a name: a skill, one of the highest-return moves a small team can make.
· Focus
The Right People in the Right Seats
Map a whole organisation's Zone of Genius and you can finally see who is in the wrong seat, and where the right answer is not another hire but an agent.
· Building
The True Cost of Vibe Coding
Vibe coding feels free, and that feeling is the trap. It is the fastest way to a first output and, past a certain point, one of the most expensive ways to run a system.
· Focus
The Fractal Zone of Genius
Zone of Genius is usually taught as a destination you arrive at. But once you spend most of your time inside it, the zone splits into a finer version of the same map. Genius is fractal.
· Focus
Real Delegation Is Deciding What You Will Never See
Human in the loop sounds responsible, but if you stay in the loop for everything you have delegated nothing. Real delegation begins where you decide to stop looking.
· The Market
AI Is the Easy Part. Knowing Where to Point It Is the Job.
Every firm can now buy the same models. The scarce thing is the judgment to know exactly where a tool creates leverage and where it creates expensive theatre.
· Method
From Chatting to Delegating
Most people underuse AI because they treat it as an oracle to consult rather than a worker to delegate to. Closing that gap is not a prompting skill but a trust climb, the same one that separates managers who can delegate from those who cannot.
· Method
The Agentic Maturity Model
Every organisation sits somewhere on a five-step climb from AI curiosity to AI autonomy. Knowing your step is worth more than any prediction about where the technology is headed.
· Method
Where AI Almost Works Is Where It Hurts You
The task AI is obviously bad at is safe, because you will never trust it. The task it does almost perfectly is the one that will hurt you, because you will.
· Method
The Four-Layer Diagnostic
Most AI projects solve the wrong problem because they accept the first one handed to them. A deliberate descent through four layers finds what the decision-maker actually wants.
· Building
Where to Draw the Line
The teams shipping reliable AI and the teams shipping liabilities both use AI. What divides them is one variable: how rigorously their outputs get verified.
· Sectors
Agentic AI for Real Estate
Real estate never had a lead problem. It had a follow-up problem, and follow-up is the one thing an agent does tirelessly, freeing your people for the moments that actually close a deal.
· Building
Context Is the Product
An AI agent's quality is mostly decided before it runs a single step, by what you put in front of it. The craft is knowing which of six kinds of context goes where.
· Sectors
Agentic AI for Wealth Managers and Family Offices
In wealth management the returns are largely commoditised. The real product is trust, and that is exactly the thing an agent cannot manufacture and the advisor has least time for.
· Focus
Return on Energy
Founders track return on capital obsessively, but the true binding constraint is return on energy: the yield on the one input you can never raise more of.
· Method
OKRs: How Google Sets Goals
OKRs are usually taught as a management ritual. The sharper truth is that a key result is the interface between your intent and the system, human or agent, that executes it. A goal you cannot measure is a goal you cannot delegate.
· Sectors
Agentic AI for Law Firms
Law is the ideal domain for AI agents and the most dangerous, for the same reason. The question is never whether AI can do the work, but who carries the liability when it is wrong.
· Building
The Maker and the Checker
The most overlooked structural decision in any agentic system is separating the agent that does the work from the agent that checks it. A model grading its own output is the most generous marker in the world.
· Sectors
Agentic AI for the SME
The enterprise has the budget; the SME has the speed. With AI, deciding on Monday and running by Friday beats a year of committees, and the smallness you thought disqualified you becomes your edge.
· Method
The Ten Jobs an AI Agent Is Actually Good At
An AI agent does not see industries, it sees jobs, and the same ten job-shapes recur in every business. Knowing which one you are drowning in, and which is safe to start with, is the whole game.
· Sectors
Agentic AI for Logistics and Supply Chain
In logistics, software already runs everything that goes right. The entire cost, and the entire case for agents, lives in the exceptions that automation was never able to touch.
· Building
Harness Engineering
A raw model is capability without a job. It can reason and draft but does not know your business or reach your systems. The harness around it is what turns that into something that ships.
· The Market
When Everyone Claims Everything, Restraint Is the Signal
The AI market is so saturated with claims that claims no longer carry information. The only thing that still signals is proof, and the discipline to claim less than you could.
· Fundamentals
Why Your AI Forgets
People are surprised a capable AI cannot remember what it did an hour ago. The surprise comes from a wrong mental model, and the fix explains most agent failures.
· Sectors
Agentic AI for Consultants and Advisory Firms
A consultant's product is judgment, but most of a consultant's hours go to the scaffolding around it. AI does not threaten the value. It threatens the cost structure.
· Sectors
Agentic AI for Banks and Financial Institutions
Banks think their control culture is why they are behind on AI. It is the exact infrastructure agentic AI requires, which makes them ahead on the part everyone else is scrambling to build.
· Focus
The Zone of Genius
Being good at something and being energised by it do not travel together. Plot them on separate axes and you find the corner where both run high, the place almost nobody spends enough of their life.
· Sectors
Agentic AI for Clinics and Hospitals
Healthcare is not behind on AI because it is cautious, but because it applies surgical caution to paperwork. Separate the clinical from the clerical, and the safe, valuable work is already in plain sight.
· Sectors
Agentic AI for the Solopreneur
AI's real gift to the solopreneur is not speed, it is staff. One person can finally be a company, but only by building the systems that do the work instead of becoming a faster bottleneck.
· Fundamentals
Chatbot, Assistant, Agent
Three words get used as if they mean the same thing. They do not, and the confusion is one of the main reasons AI projects disappoint the people who commissioned them.
· Sectors
Agentic AI for Marketing and Creative Agencies
AI did not commoditise creative. It made the average free and pushed the whole business above the line, into the taste and judgment a client cannot prompt for themselves.
· Fundamentals
The Model Is the Engine. The System Is the Car, the Road, and the Traffic Laws.
Almost everyone watching AI tracks the smartest model of the month. It is the most visible part of the field and the least decisive. Swapping engines does not build you a better car.
· Fundamentals
What Is an AI Agent, Really?
An AI agent is software that pursues a goal through a loop of decisions and actions. It is not a chatbot with a confident personality, and that difference is everything.
Work with CLRT
Find the work worth giving to an agent.
Begin with the workflow, the business goal and the judgement that must stay human.