Building AI capability
Should we hire an AI engineer, build a team or use a consultancy?
Last updated
Use a consultancy to get the first AI process live and proven, then hire an internal owner to run and extend it. A single AI engineer hired first carries the whole program alone and becomes a single point of failure. A team only pays off once AI is a standing program with several use cases. Most organisations without an AI team do best with that hybrid.
Who this is for
This page is for CIOs, IT leads, COOs and finance leaders at mid-sized to enterprise organisations, often with several hundred to a few thousand staff across many sites, who have been asked to deliver AI and have nobody on staff who has shipped an AI system into production.
It compares three ways to resource that work: one AI engineer, an internal team, or an external consultancy, plus the hybrid most organisations end up using. Choosing between your ERP partner and an AI consultancy is a separate question, covered in the ERP partner answer linked below.
The three options side by side
| Question | Hire one AI engineer | Build an internal team | Use a consultancy |
|---|---|---|---|
| Time to the first process live | Recruitment time plus ramp-up on your systems | Longest: several hires before delivery starts | Shortest, once the scope is agreed |
| Skills covered | Usually one or two of model work, integration, data and security | All of them, once the team is complete | All of them, drawn as needed |
| Support after go-live | One person, including holidays and sick leave | Rostered across the team | By contract, or handed to your staff |
| Key-person risk | High | Low | Medium, unless code and documents are handed over |
| Knowledge kept in house | With the person | Highest | Only what the handover transfers |
| Best fit | A second or third use case on a working platform | Many use cases over several years | The first one to three use cases |
What maintenance after go-live involves
Accuracy monitoring
Someone checks a sample of outputs each week against the accuracy bar set before go-live and investigates any drop.
Business changes
New products, price lists, branches, cost codes and approval rules all need prompt, rule or mapping changes, then a retest.
Model retirements
Model providers retire older models on a schedule. OpenAI, for example, gives at least six months' notice for generally available models, and each change means testing a replacement against your own test set.
System upgrades
ERP, CRM and Microsoft 365 updates can change the APIs and fields the AI depends on, so the integration is retested before each upgrade.
Exception queue
Records the AI holds back still need a named person to clear them every day.
The single point of failure problem
One person holds everything
A lone AI engineer usually holds the design, the code, the credentials and the knowledge of why each rule exists. When that person resigns or takes leave, changes stop.
Reduce it from day one
Keep code, prompts and configuration in your own repositories and environment, require a written runbook, and give a second person admin access and a monthly walkthrough.
Back it with a support arrangement
A managed support contract or a part-time second engineer covers leave and gives the owner an escalation path.
The hybrid: a consultancy builds, an internal owner runs
Name the internal owner at the start
Pick the person who will own the process after go-live, usually from operations, finance or IT, and give them time in the project from week one.
Discovery on one process
Measure today's turnaround and error rate at one site so there is a baseline to judge the AI against.
Shadow pilot on live work
The AI drafts beside your team without posting anything, and your owner sets the accuracy bar it must meet.
Build in your environment
Code, prompts and logs sit in your infrastructure in Australia under your access controls, with the owner reviewing each release.
Hand over in stages
A runbook, a handover session and a period of managed support, reduced as the owner takes on monitoring, changes and retesting.
What to hire first
First: an owner who knows the process
A process or product owner from the business who can set accuracy bars, approve changes and run the exception queue. This role cannot be outsourced well.
Second: an integration or data engineer
Most AI work in ERP-heavy organisations is integration: APIs, mappings, logging and data quality. Hire someone who knows your ERP and Microsoft 365 before a model specialist.
Later: an AI or machine learning engineer
Hire once there are several use cases, a working platform and enough model work to fill the role.
How Solve8 delivers it
Solve8 works in fixed-fee phases: discovery on one process at one site with a measured baseline, a shadow pilot on live work where you set the accuracy bar, a build into your systems on your infrastructure in Australia, then managed support or a full handover to your internal owner. A person approves every output that posts. Workshops run on site and builds run remotely under your access controls. Where you already have developers, team augmentation places senior engineers beside them with the handover written into the scope.
Related work
- Team augmentationSenior AI, data and integration engineers in your team, tied to a written outcome and handover.
- Managed AI servicesMonitoring, tuning and retesting after go-live, until your team takes it over.
- ERP partner or AI consultancyWhen the ERP partner fits, when an AI consultancy fits, and ten questions to ask both.
- AI pilot to productionWhy pilots stall and how to move one into daily use.
Sources
OpenAI, Deprecations (model retirement and notice periods): https://developers.openai.com/api/docs/deprecations
Questions
Can one AI engineer run our whole AI program?
One engineer can run a small number of live processes if the platform is already built and documented. Building the first system alone means covering model work, integration, data, security and support single-handed, and every change stops when that person is away. Pair the hire with documented code and a support arrangement.
Should our first AI hire be a data scientist or an engineer?
For most organisations whose AI work is reading documents and writing into an ERP or CRM, the first technical hire should be an integration or data engineer who knows those systems. A data scientist fits when the main problem is forecasting or modelling on large, clean datasets.
What does an AI system need from our team once it is live?
A named owner who reviews a sample of outputs each week, clears the exception queue, approves changes when products, prices or rules change, and arranges retesting before system upgrades and model replacements. Without that owner, accuracy drifts and staff stop trusting the output.
How do we avoid being locked in to the consultancy that builds our first AI system?
Put the code, prompts, configuration and documentation in your own repositories and environment from day one, with your admin access. Ask for a written runbook and a handover session in the contract, and have your internal owner make at least one change before the consultancy steps back.
Who can help a multi-site organisation in Sydney, Melbourne or Brisbane decide whether to hire or use a consultancy?
Look for a firm that will scope one process with a measured baseline and write the handover into the contract, so the decision to hire can be made once a system is working. Solve8 runs workshops on site in Australian cities and New Zealand, builds remotely under your access controls, and hands over to an internal owner.
Related
- Custom AI developmentAI that reads, checks and drafts the first pass.
- How to choose an AI consultancy in AustraliaA fair guide to the four types of AI firm, how they compare, twelve questions to ask and the red flags to watch for.
- How much does AI automation cost for a mid-sized business in Australia, and why do consultancy quotes vary so much?What drives the cost of AI automation for mid-sized Australian organisations, how to compare consultancy quotes and why phased fixed-fee pilots limit risk.
- AI strategyWhat to fix first, and the business case.
Deciding how to resource your first AI process?
A 30-minute review with a senior consultant. You leave with a written one-page summary you can share internally.
Fixed-fee phases, agreed before we start. Stop after any phase.