AI Agents for Business: Always-On Automation Explained

Your AI agent for business just started working without you.

That’s the shift OpenAI announced at DevDay. Dots—always-on agents that run in the background and execute multi-step work without human prompts—signals a hard pivot from “chatbot you talk to” to “agent that works for you while you sleep.”

For finance teams, supply chain managers, and operations leaders, this is not a nice-to-have. This is how enterprise work changes in 2026.

---

The Three Things That Changed

1. Work no longer stops when you leave the desk

Until now, AI required human input. You ask. You wait. You act on the answer. Dots reverses that. You set a goal (reconcile yesterday’s invoices, monitor our data pipeline, gather competitor news) and the agent runs it asynchronously, reporting back when done. It operates across 4,000+ apps—email, Salesforce, your accounting system, your data warehouse.

For organisations running daily cycles, this means: 8-hour work compresses to overnight execution. A finance team that spends three hours every morning on reconciliation frees those three hours. AI agents for business make overnight execution standard.

2. You pay for execution now, not just access

OpenAI introduced GPT-6.1 Sol: a new model that matches the raw smarts of their flagship tier (GPT-6 Astra) at one-fifth the cost. Astra is for high-stakes reasoning. Sol is for high-volume, background work.

Here’s why this matters: Running AI agents for business 24/7 on frontier-tier models is ruinously expensive. Sol makes autonomous workflows economically feasible. You run supply chain agents on Sol. You run strategic planning on Astra.

The pricing tiers reflect this. The new $500/month tier is for teams running persistent agent fleets with guaranteed compute. Standard tiers are for interactive, human-driven work.

What this means: Your IT cost for AI is no longer “per seat.” It’s “per workflow, per compute tier, per execution volume.” Budget differently.

3. Governance moves from access control to agent permissions

When AI was a chatbot, governance was simple: who can log in? Now it’s harder. Agents make decisions and take actions. An agent that reconciles invoices needs read and write access to your accounting system. An agent monitoring customer data needs those credentials without human intervention.

This is not permission management. This is entrusting software with your keys.

Organisations that move fast here will define strict boundary lines: what data can this agent touch? What actions can it take? When does it ask before acting? What audit trail do we keep?

---

Three Immediate Steps

Step 1: Map your overnight and recurring work

List processes that run today but no one is actively steering—daily reconciliation, report generation, data cleanup, competitor monitoring, invoice matching. These are AI agent candidates for business.

Why: Agents excel at unattended, multi-step routines. A CFO spending four hours on month-end close is not an agent candidate. A billing team reconciling yesterday’s transactions at 9 a.m. every day is.

Step 2: Lock down agent permissions before you run them

Define exactly what each agent can read, write, and approve. Build an audit log so you see what every agent did and when. Set approval gates—if an agent flags a reconciliation discrepancy above £50k, it asks a human before acting, not after.

Why: Agents operate without supervision. Permissions are your only control.

Step 3: Pick the right model for each task

Routine, high-volume background work runs on Sol (cost-effective). Strategic decisions, novel problems, complex reasoning run on Astra (maximum intelligence). Mixing them up wastes either money or brainpower.

Why: Your total cost depends on which model does which job.

---

Who This Affects Now

Finance teams benefit first: reconciliation, invoice matching, expense approval loops.

Supply chain comes next: order status tracking, supplier alerts, inventory updates.

Anywhere with recurring, multi-step, data-driven work is fair game for AI agents in business.

If you’re running executive AI workshops or designing AI capability for your team, this reshapes your timeline and your cost model.

---

The Bottom Line

AI agents for business make overnight work possible. That changes your cost structure, your governance model, and where you focus your team’s time.

Leaders who map this clearly—which processes go autonomous, how tightly you lock permissions, which model tier you use—will ship this in weeks. Leaders who don’t will cost themselves months and misallocate budget.

I’ve reviewed 1,000+ AI projects in real organisations. The ones moving fastest on AI agents right now are not waiting for perfection. They’re identifying one good candidate, building tight permissions, and shipping it. Then they do the next one.

Start this week. Pick one overnight process your team owns. Map what it needs. Set the permissions. Ship it.

That’s how you move from “AI is interesting” to “AI is how we work now.”

Ready to build this into your organisation? Speak to me about keynotes or a custom workshop to make agents real for your leadership team.

---

Frequently Asked Questions

Do I need the $500 ChatGPT tier to use AI agents for business?

No. Dots and background agent work roll out first to ChatGPT Pro, Business Premium, and Enterprise tiers. You don’t need the top tier. You do need the right tier for your compute volume.

Are AI agents for business secure for regulated industries (finance, healthcare)?

Agents are as secure as the permissions you give them and the audit trail you keep. Healthcare and finance teams should start with low-stakes workflows—not customer data, not financial decisions—until governance is rock solid. Then graduate to higher-stakes work once you have the controls in place.

What if an agent makes a mistake?

It depends on the mistake and whether you built human approval gates. If an agent mismatches an invoice and you have a £50k approval gate, a human catches it. If you don’t, you discover it later in the month-end close. Start with approval gates on high-value actions.

How long until this is mainstream?

AI agents are available now to paid subscribers. Wider adoption in organisations happens when IT and Finance align on permissions and cost structures. That’s a 2–6 month conversation in most places.

Should I retrain my team now?

No. Retraining on “how to use agents” is premature. What you need now is clarity on which processes to automate and how to govern them. The “how to use the agent” part is trivial once those decisions are made.