Where business AI may go next: facts, forecasts and open questions

Why this page sorts claims into three piles

Every week brings a new statement about what AI will do to business. Some of it is true today. Some is a vendor's roadmap. Some is a guess dressed as a certainty. Treating them the same way leads to two mistakes: paying for things that do not exist yet, and ignoring things that already work.

So this page keeps three piles. Demonstrated means a vendor documents the capability, you can use it, and the limits are written down. Announced means a vendor or a standards group has said it is coming or partly delivered. Scenario means a plausible future that nobody has shown at scale. We date everything and we say which pile it sits in. The page is reviewed every two months because the piles move.

Demonstrated: what businesses can use now

Start with a survey number, carefully. Statistics Canada reported in June 2026 that 19.2% of Canadian businesses used AI to produce goods or deliver services over the twelve months before the survey, and 9.2% in construction [S07]. That is a defined measure of a specific question over a specific period. It does not count informal use of a chat assistant, and it says nothing about what those businesses got from it. Use it as a marker that adoption is real and uneven, not as a trend line.

On capability, the vendor documentation is more useful than the headlines because it states limits. OpenAI's agents documentation describes agents as applications that plan, call tools, collaborate across specialists and keep enough state to finish multi-step work, and it documents guardrails and approval flows that pause before risky steps [S29]. Anthropic's computer-use documentation describes a model that can take screenshots and operate a mouse and keyboard, alongside firm advice to run it in an isolated environment, restrict its network access and have a person confirm decisions with real-world consequences [S30]. Google Cloud documents grounding models in your own documents and search results, which is covered on the documents page.

• Reading and answering from your documents with citations back to the passage. Demonstrated, with the freshness and permissions problems covered in how AI uses your documents.
• Calling tools and taking actions in software through defined connections. Demonstrated. The action is only as safe as the authority you gave it, which is the subject of approval boundaries.
• Handing work between specialised steps and pausing for a person. Demonstrated; the patterns are described publicly [S11, S29].
• Operating a desktop like a person would. Demonstrated within limits the vendor spells out, including that the model may follow instructions it finds on screen [S30].

Notice what every one of those has in common: the capability exists, and the vendor tells you to keep a person in the loop for anything that matters. That is not marketing caution. It is the current state of the tools.

Announced and partly delivered: the connective tissue

The gap between one assistant doing one task and a system carrying work across a business is mostly plumbing. Three pieces of plumbing are moving right now.

Interoperability standards. The Model Context Protocol is described by its maintainers as an open-source standard for connecting AI applications to external systems, and its site lists support from multiple assistants and developer tools [S34]. That matters for a business because it lowers the cost of connecting an assistant to a calendar, a database or a document store without a custom build each time. How completely vendors adopt it, and how well it holds up for permissions and audit, is still being worked out.

Agents that span many systems. Vendor documentation now describes multi-agent handoffs and state kept across steps [S29]. What is announced rather than proven at scale is agents reliably carrying a whole business process across several companies' tools, over weeks, with the exceptions handled. Pieces of that exist. The whole is not something you can buy off the shelf and trust unattended.

Managing agents like staff. Approval flows, logs and guardrails are documented [S29, S30]. What is still forming is the everyday management layer: who owns an agent, how its authority is reviewed, how you find out it failed quietly. Expect this to be where most of the practical progress happens over the next year, because it is the part businesses actually ask for.

Scenario: what nobody has shown yet

These are the claims to treat as predictions. They may come true. Nobody can honestly put a date on them, and we will not.

• AI runs most of a working day for an ordinary business. Requires the plumbing above plus records, authority and exception handling that most businesses have not designed. The companion article walks through the conditions.
• Businesses reorganise around agents rather than departments. Plausible in software-heavy firms first. No evidence yet for trades and service businesses at scale.
• Large-scale job replacement. A prediction with a wide range of outcomes. We do not publish job-loss claims, and we would treat anyone who states a number with a date as guessing.
• Models become reliable enough to skip human checks. Vendor guidance today says the opposite for consequential actions [S30]. Until that guidance changes, treat the claim as a scenario.

What this means for an owner

The tempting response is to wait until the picture is clear. The problem is that the things that decide how much AI can carry later are things you control now, and they take time. An assistant can only act on records that exist. It can only respect authority that has been written down. It can only hand work between tools that know about each other. None of that depends on which model wins.

1. 01Get the records in orderOne authoritative version of each document, an owner per record, a next step and date on every open item. The handoff assessment finds where this is missing.
2. 02Write down authorityWhat may be done without asking, what needs approval, who can revoke it. This is the approval boundaries worksheet, and it is useful whether or not you ever buy an agent.
3. 03Run one bounded experimentA single recurring task, a person who owns it, a log, a stop rule. The what to automate first guide sets it up.
4. 04Keep a watch sheetWrite the scenarios you are planning around and what evidence would change your mind. Review it on a date. The sheet below is for that.

Questions people ask

What can AI agents demonstrably do for businesses today?

Read and answer from documents with citations, call tools and take defined actions in software, hand work between specialised steps, operate a desktop within stated limits, and pause for a person's approval. Vendors document each of these along with advice to keep a person confirming consequential actions.

Which claims about the future of AI are still predictions?

That AI will run most of an ordinary working day, that businesses will reorganise around agents, that human checks will become unnecessary, and any job-displacement figure with a date on it. Plausible, undated, unproven at scale.

Will AI run my business operations for me?

Not unattended, not today, and not without records, authority and connections that most businesses have not built. What it can do now is carry defined work between the tools you have, with you approving what matters.

How often does this page change?

It is reviewed every two months. The last-checked date at the top is real. If a claim moves piles, the table changes and the update date moves with it.