On 12 November 2026, Xero announced Xforce, a beta programme that places forward-deployed engineers and partner consultants inside accounting firms to redesign the firm's workflows around AI, with Australian firm Liston Newton Advisory named as a design partner. Strip away the branding and what Xero has published is an admission with real commercial weight behind it: shipping AI features into a product does not, on its own, change how a firm works. Someone has to go into the practice, look at how a job actually moves from client email to lodged return, and rebuild that path. A software vendor is now funding people to do that, which tells you where the value has moved.
For Australian accounting, bookkeeping and advisory firms, that is a useful signal regardless of whether you ever join the beta. It says the binding constraint on getting value out of AI is no longer model capability or token cost. It is process design — knowing which steps in your workflow are decisions, which are transcription, which are chasing, and which exist only because the software could not do them before.
What "forward-deployed" actually means in practice
The term comes out of enterprise software, where vendors send engineers to sit with a customer and build on top of the platform rather than hand over documentation and hope. In an accounting practice, a forward-deployed engagement looks less like a software rollout and more like a time-and-motion study with a build attached.
The sequence is roughly: watch a real job run end to end, instrument where time actually goes, pick one workflow that is high-volume and low-judgement, rebuild it with AI handling the mechanical parts, measure, then move to the next. The engineer is not there to configure a product. They are there to decide what the product should be pointed at, and to write the connective code between the ledger, the document store, the email system and the practice management tool that nobody's roadmap covers.
Two things follow from that model that firms consistently underestimate. The first is that most of the work is integration, not AI. The model that reads a supplier invoice is the easy part; getting that invoice from a shared mailbox, through an approval, into the correct ledger account, with an audit trail that survives a review, is where the effort sits. The second is that the engagement produces a changed process, not a tool. If the firm reverts to the old process the moment the engineer leaves, nothing has been bought.
The four accounting workflows that are realistic candidates
Not everything in a practice is worth redesigning. These four come up because they are high-volume, repetitive, and have a clear definition of done.
Transaction coding and bank reconciliation
The highest-volume candidate in almost any firm. AI classification handles the routine lines, confidence thresholds route the ambiguous ones to a human, and the reviewer sees only exceptions. The design question is not whether a model can code a transaction — it can — but what confidence threshold you set, who reviews below it, and how corrections feed back so the same supplier is not misclassified every month. Get the feedback loop wrong and you have built a system that generates the same work twice.
Client document chasing and intake
Every firm loses hours to the gap between "we need your logbook" and receiving it. An automated intake workflow tracks what is outstanding per client, drafts the follow-up, escalates on a schedule, reads what comes back, classifies it, and files it against the right job. The AI component is modest. The value is in the state tracking — knowing, at any moment, exactly what is missing across 300 clients without anyone maintaining a spreadsheet.
Workflow triage and job allocation
Inbound work arrives as email, portal messages, phone notes and forms. Classifying each item by urgency, client, job type and required skill, then routing it to the right person with the relevant context attached, is a well-bounded AI task. It is also where firms get the fastest felt improvement, because it removes the daily triage meeting rather than shaving minutes off a task.
Advisory reporting and first-draft commentary
Management reporting commentary is the one with the most judgement in it, and therefore the one to attempt last. A model can produce a competent first draft of variance commentary from the ledger and prior-period data. It cannot decide what matters to that client this quarter. The honest framing is that it compresses drafting time for a senior who still writes the conclusion — which is worth having, and is not the same as automation.
Firms that work on client files under professional obligations face a specific version of this problem, and designing AI into professional services workflows is less about the model than about where the human sign-off sits and whether the audit trail would survive scrutiny.
What a firm needs in place before anyone deploys into its stack
A forward-deployed engagement fails quietly when the engineer spends the first fortnight discovering that the firm's data is not in the state anyone assumed. Four things are worth having sorted before the work starts.
A written description of one workflow, end to end. Not a policy document — the actual steps, including the undocumented ones where someone checks a thing because of what happened in 2023. If you cannot describe it, it cannot be redesigned.
Clean system boundaries. Which system is the source of truth for client contact details, for job status, for document storage. Where two systems both claim to be authoritative, the automation will pick one and quietly diverge from the other.
A decision on data residency. Australian accounting firms hold tax file numbers, bank data and personal financial information. Whether that data may leave Australia, which sub-processors are acceptable, and whether any of it needs to stay inside infrastructure you control, are decisions to make before a pilot, not after. For firms where client agreements or risk appetite rule out offshore processing, Australian-hosted or self-hosted deployment is the path, and it changes the architecture enough that retrofitting it later is painful.
A named internal owner. One person who can approve a process change without a partner meeting. Engagements that need consensus for every change produce demos, not deployments.
The security responsibilities the vendor will not own
The failure mode in agent deployments is rarely the model. It is the infrastructure around it. A security incident log reports that on 8 November 2026 a NEAR AI inference server running a deprecated Private Chat deployment was compromised, with data from 18 agents, roughly 100 user accounts, and API keys belonging to two cloud customers exposed in logs. That account comes from an incident log rather than a primary vendor post-mortem, so treat it as reported rather than independently confirmed — but the shape of it is instructive.
Three things stand out. A deprecated deployment was still running and still reachable. Credentials were sitting in logs. And agent state and prompts — which in an accounting context would contain client financial detail — were in the same exposure as the keys.
Those are all customer-side responsibilities in almost every vendor agreement. A forward-deployed engineer will build you a workflow. They will not, as a matter of course, own your secret rotation, your log redaction policy, your decommissioning of last year's pilot, or the question of which staff can invoke an agent that has write access to the ledger. If you run agents, assume prompts and logs are as sensitive as the records they touch, and treat retired pilots as live risk until they are switched off and their keys revoked.
Why the economics changed this year
The reason vendors are now spending money on implementation people is that the model layer stopped being the expensive part. November 2026 pricing trackers list models spanning from Muse Ember at around US$0.10 per million input tokens to GPT-6-Cyber at US$20/US$100, a range of two orders of magnitude for tasks that overlap considerably. Anthropic's Claude Fable 5.2, released on 9 November 2026 at US$8 and US$40 per million input and output tokens, came in roughly 20% below its predecessor. At roughly AUD 0.15 per million input tokens for the cheapest tier, classifying a year of a client's transactions costs less than the coffee consumed deciding whether to do it.
What has not fallen is the cost of knowing which step to automate and wiring it into systems that were never designed to talk to each other. That is the work Xero is now paying engineers to do inside firms, and it is the work that determines whether any of this shows up in realised hours.
A note on budgeting: several current rates are promotional. Trackers report Google pricing scheduled to change on 1 December 2026 and some Gemini Flash introductory rates holding only until 31 December 2026. Those figures come from pricing trackers rather than vendor pages, so confirm them directly — but design for model substitution regardless, so a rate change is a config edit rather than a rebuild.
The first move is to pick one workflow and time it honestly for a fortnight. Count how many client documents arrive, how many touches each one takes before it is filed, and how many of those touches involve a judgement rather than a transcription. That ratio tells you whether the workflow is worth redesigning before anyone quotes you for it.
If you want the next step done properly, we will run an assessment across your practice stack and come back with a scoped build for one workflow — what it integrates with, where the data sits, whether it can run Australian-hosted, a fixed delivery window that is usually two to four weeks, and a straight answer if the honest conclusion is that the workflow is not worth automating yet. Book an assessment and we will tell you which of the four candidates above is the one for your firm.