Table of Contents
Agentic AI has become one of those terms that means whatever the person using it needs it to mean. Zaid Aboobaker, founder and CEO of CompassPoint Consulting, a fractional CFO and corporate finance firm operating across the UAE and UK, doesn’t have much patience for that version of it. He draws a hard line between a chatbot bolted onto rules-based automation and a system that can actually reason through an exception, and he’s built his own firm’s month-end close around the difference: three days of work down to about four hours.
We asked him where agentic AI actually earns its place in a finance function, where he draws the line on autonomous action, and about a specific incident where an agentic workflow got something wrong in front of his own investors.
In conversation with CompassPoint’s CEO Zaid Aboobaker
Agentic AI is a loaded term right now, most of what gets called ‘agentic’ in finance is still rules-based automation with a chatbot layer on top. What does an actual agentic system do inside a finance function that RPA or a traditional automation tool couldn’t, and where have you seen that distinction actually matter for a client?
Zaid Aboobaker: Traditional automation follows a fixed path. If the reference matches, reconcile. If it does not, stop and escalate to a human. It cannot reason about the exception. It can only route it.
An agentic system reasons about the goal. Faced with a payment that matches no single invoice, it can work out that the amount equals three invoices combined, less the early-settlement discount that client always takes, and resolve it. It chooses the steps rather than following a fixed route.
Where it matters for us is at the level of a whole workflow, not a single task. We run agentic close, GL reconciliation and a weekly 13-week cash forecast across our own UAE and UK entities and our clients’ books. Our own month-end close and board reporting went from three days of work to about four hours. On one client engagement, we cut the monthly close cycle by a full week. A rules-based tool could not have done that, because the work is full of exceptions, and exceptions are exactly where traditional automation stops and agentic systems keep going.
CompassPoint runs a fractional CFO model, which means the same team is covering multiple clients’ books at once. How does agentic AI change what a fractional CFO can realistically take on?
Zaid Aboobaker: The lazy read is that agentic AI lets you serve more clients with the same headcount. True at the margin, but if that is all you use it for, you are just running the same commoditised service slightly cheaper. That is not our model.
The real change is the mix of work per client. Deliverables that used to consume senior time, such as board packs, cash forecasts, close checklists and reconciliations, now run on agentic workflows with a human reviewing the output. Our own board reporting cycle dropped from three days to about four hours. That senior time does not disappear. It moves into interpreting the numbers rather than assembling them. More scenario work, more time on the actual decision.
The commercial effect is that a founder running a business of 15 to 100 people, who could never justify a $200,000 full-time CFO, now gets CFO-grade forecasting and board-ready reporting at a fraction of that. So it is not really about capacity. It is about moving up the value chain for each client. The processing gets cheaper and faster. The judgement becomes the product.
Financial reporting is a trust-and-liability business; a wrong number has real consequences. Where do you draw the line on autonomous action versus human sign-off, and has that line moved in the last year?
Zaid Aboobaker: The line is drawn by consequence and reversibility, not by capability, and that principle has not moved.
If an action is reversible, low-consequence and auditable, an agent can do it autonomously: categorising transactions, preparing reconciliations for review, drafting commentary, flagging anomalies. If it is irreversible or high-consequence, a human signs off before it happens: releasing a payment, anything touching a statutory filing, anything that becomes the board-facing record.
The rule we run to is simple. Every number that leaves our function has a named human accountable for it. The agent can assemble an entire 32-slide board pack end to end, but a person signs it off before it reaches a client or a board.
What has changed in the last year is the size of the autonomous zone. The tools are more reliable, and the audit trails are better, so more sits safely on the autonomous side than twelve months ago. But the test is unchanged. The fact that an agent can do something is not a reason to let it.
What is the actual failure mode you worry about most?
Zaid Aboobaker: The failure I design against is the plausible, completed, wrong result. The task finishes, the output looks right, nothing throws an error, because technically there was none. The agent simply applied the wrong logic with full confidence.
In reconciliation, the specific risk is losing the exception. A traditional tool escalates what it cannot match, and that queue is where a human catches problems. An agent that resolves those items can remove the signal the human used to rely on. The work looks more complete when it deserves more scrutiny. The moment a reconciliation shows as balanced and closed, the instinct is to trust it.
That is the entire reason we keep a named human accountable for every number rather than letting the agent publish itself. Our controls exist for this failure specifically. Remember, our clients make a very personal decision when they choose to entrust us with their financials. If we send things that do not reconcile, our credibility is shot to bits in seconds.
I had an incident where an agentic workflow prepared a cash flow forecast for my investors. I had given them a heads-up that we were testing, and the result was a double-counted item leading to an inflated cash balance. It was proof enough that, while everything can look right, it’s easy for the agent to make an assumption that gets buried in a model, creating an inaccuracy.
How does adoption of this look different for finance teams in the GCC compared to what you’d see in the US or Europe, is it moving faster, slower, or just differently, and what’s actually driving that?
Zaid Aboobaker: The UAE has set a target to deliver half of government services through AI by 2028, and the government leading by example changes what the private sector treats as normal. Many businesses here are young enough that they are not unwinding decades of entrenched systems the way a European bank is. The UAE has a leadership culture that’s comfortable with moving quickly.
In North America and Europe, businesses are more mature, and simply bolting AI into an existing process is a recipe for disaster. Therefore, we see more large-scale, multifaceted transformation programs that impact not just finance but other areas of the business. It is also important to remember that the GCC’s regulatory landscape is continuing to evolve. The speed at which AI can be adopted by banks and financial institutions is greater because the rules are being written around it to some extent.
For a finance leader who has not touched agentic AI yet, what should they do first, and what mistake should they avoid?
Zaid Aboobaker: Try to work with one AI provider. Many people bounce between Copilot, ChatGPT, Gemini, Claude and many others. It is better to pick one, get a subscription that protects your data, and allows you to turn off the functionality that lets the model learn from your information. Then build up your context and allow the AI to memorise key facts.
Then pick one low-consequence process and run the agent alongside your existing process, not instead of it. Let it operate in parallel and compare its output to what your team would have produced. You learn where it is reliable and where it is not before anything depends on it. Trust is earned in parallel. It is roughly the path we took ourselves, from off-the-shelf tools to our own agents to a full operating system, over about two years, not two weeks.
The mistake to avoid is pulling the human review too early. The danger is when it has been right often enough that people quietly stop checking. Avoid this temptation. That is exactly when a confident, plausible, wrong result gets through. Do not confuse ‘works most of the time’ with ‘can be trusted unsupervised.’
Editor’s take: The line worth sitting with is ‘the fact an agent can do something is not the reason to let it.’ Most agentic AI commentary sells capability. Aboobaker sells the governance around it instead and backs it with a real admission: a double-counted cash-flow item that reached his own investors during a test. That’s not a hypothetical risk he’s warning other people about; it’s something that actually happened to him, and he named it. Worth watching as CompassPoint scales this further: the time-saving figures here, three days to four hours, a week off a client’s close, are self-reported by the firm running the workflow, not independently audited. That doesn’t make them wrong, but it’s the kind of claim that gets stronger with a named client willing to confirm it on the record.