The short version
An Australian asset manager with 20 users wants an AI assistant that works across its data. Buying it from a US vendor costs roughly $486,000 in the first year. Buying it from us costs roughly $160,000. Building it from scratch costs more than either.
The gap has almost nothing to do with AI. It is about who does the work — and whether anyone checks that it worked.
The comparison
Every line below is a real cost someone pays. First year, 20 users, same scope.
| Section | US vendor | Build it yourself | Datafabric |
|---|---|---|---|
| AI platformEverything needed to serve the models: the assistant licences and usage fees, plus routing, guardrails, a vector store, evaluation and observability. Salesforce bundles most of it into Agentforce. Build it yourself and you license and host each piece.i Everything needed to serve the models: the assistant licences and usage fees, plus routing, guardrails, a vector store, evaluation and observability. Salesforce bundles most of it into Agentforce. Build it yourself and you license and host each piece. |
$8,000Agentforce consumption at roughly $2 a conversation, which is about $2,000 at this volume and largely covered by the Foundations free tier, plus roughly $6,000 of Claude or ChatGPT seats for the team.i | $35,000$6,000 assistant seats, $6,000 vector store at Pinecone’s $500 a month enterprise tier, $12,000 of observability and evaluation seats, and $10,000 to host the gateway and guardrails.i | Included |
| Data platformWhere your data is stored, joined and modelled — the warehouse plus the ingestion, transformation and orchestration around it.i Where your data is stored, joined and modelled — the warehouse plus the ingestion, transformation and orchestration around it. |
$60,000Salesforce Data 360 Starter at list price, which includes 10 million data credits and 5TB of storage.i | $60,000A warehouse at the $28,000 to $50,000 benchmark for this size, plus ingestion, transformation and orchestration tooling on top.i | Included |
| BI platformDashboards and reporting over that data. A US vendor sells Tableau seats; on your own, Power BI is markedly cheaper at this headcount.i Dashboards and reporting over that data. A US vendor sells Tableau seats; on your own, Power BI is markedly cheaper at this headcount. |
$10,000Tableau Cloud Standard: 5 Creators at $75 a month, 10 Explorers at $42 and 5 Viewers at $15 comes to $10,440 a year.i | $6,000Power BI Premium Per User at $24 a month for 20 users is $5,760 a year.i | Included |
| Compliance auditsIndependent review of controls, access and data handling. No vendor can discharge this for you, ourselves included — you remain accountable for assets you do not operate. What we include is the programme and the evidence behind it, run once across every client. The sign-off stays yours.i Independent review of controls, access and data handling. No vendor can discharge this for you, ourselves included — you remain accountable for assets you do not operate. What we include is the programme and the evidence behind it, run once across every client. The sign-off stays yours. |
$8,000A reduced programme. The vendor’s certifications answer the infrastructure questions, but your own controls, access and data handling are still audited.i | $20,000A full assurance programme. The figure excludes the internal engineering time spent answering it.i | Included |
| People to run itDay-to-day administration: users, permissions, data quality, and keeping the feeds and integrations healthy. Roughly one full-time person. With us it is part of the service.i Day-to-day administration: users, permissions, data quality, and keeping the feeds and integrations healthy. Roughly one full-time person. With us it is part of the service. |
$200,000One administrator, fully loaded: a $150,000 salary plus about 30% for superannuation, payroll tax, workers’ compensation and equipment.i | $300,000An engineer and a half. Administering a vendor’s SaaS is configuration; running your own stack adds infrastructure, upgrades, dependency management, security patching and on-call.i | Included |
| People to build Industry Best Practice Infrastructure, Architecture, Governance, Tools, plugins and know how (one-off)Building the equivalent of what our platform already provides — standing up the infrastructure, the architecture, the governance, the feeds, the plugins and the accumulated know-how. The platform rows above are licences; this is the work. A one-off project for the other two options. With us it already exists.i Building the equivalent of what our platform already provides — standing up the infrastructure, the architecture, the governance, the feeds, the plugins and the accumulated know-how. The platform rows above are licences; this is the work. A one-off project for the other two options. With us it already exists. |
$200,000About one engineer-year at $195,000 fully loaded. The vendor supplies the infrastructure, the agent framework and the governance scaffolding, so what is left is the six data feeds, the configuration and the handover.i | $400,000About two engineer-years. With no agent framework, governance scaffolding or prebuilt connectors to start from, every piece is written from scratch.i | Included |
| Total, year one | $486,000 | $821,000 | $160,000 |
| Total, following years | $286,000 | $421,000 | $160,000 |
That is 67% and 81% less in the first year — and 44% and 62% less in every year after it.
One thing the table cannot price: whether any of it works. All three options can stand up an assistant. Only one of them tells you, six months later, whether it changed anything — which conversations it generated, which flows it influenced, which of your assumptions turned out to be wrong. Salesforce’s Testing Center reports whether an agent answered correctly, which is not the same question. Build it yourself and measurement is one more thing to build.
The second total is the one to plan against. Building the architecture, governance and tooling is a one-off, so it drops out after year one — which is why the other two columns fall and ours does not move. What never drops out is the person running it.
What we’re pricing
All three options have to deliver the same four things:
- A data platform that holds CRM data and the data that is not in the CRM.
- An AI assistant that can answer questions across all of it.
- The connections that pull that data in.
- Someone to keep the whole thing running.
We use the first year because that is the number a budget gets approved against.
One cost is left out of all three: the CRM licence itself. Most firms already pay for Salesforce and will keep paying for it whichever option they choose, so it cancels out.
Two terms worth defining, because they do most of the damage to budgets:
- Fully loaded cost — what an employee really costs. A $150,000 salary plus 12% superannuation, payroll tax, workers’ compensation, equipment and software comes to about $195,000.
- Unit registry and platform files — the data showing which investment platforms your fund flows came through, and which advisers drove them. It lives outside your CRM.
Option 1: a US vendor stack
Add Salesforce’s AI product, Agentforce, to the CRM you already own. Add their data platform, Data Cloud, underneath it.
The AI platform is cheap. Twenty seats of Claude Team or ChatGPT Business cost about $6,000 a year, and Salesforce prices its Agentforce agents per conversation at roughly $2 each, which comes to a similar figure at normal volumes. The agent machinery itself — routing, guardrails, evaluation — Salesforce bundles in. Salesforce Foundations goes further and includes 200,000 Flex Credits, Agent Builder and Prompt Builder at no cost for Enterprise Edition customers, so at low volume the agent layer really is close to free. Call the whole layer $8,000, most of it assistant seats.
It is the cheapest line on the page — but it is the one that grows. At $2 a conversation, a team having 14,000 conversations a year pays $28,000 for the same thing. Metering charges you more precisely as adoption improves.
The data platform is not. The entry-level Data Cloud package lists at $60,000 a year. Foundations includes 250,000 Data Cloud credits free, which may carry a small deployment for a while — but credits are consumed by ingestion and processing, and six platform and registry feeds consume them quickly. Reporting on top is separate again: Tableau or CRM Analytics seats come to around $10,000 for a team this size.
The data feeds are the surprise. Salesforce ships connectors for Snowflake, Databricks, BigQuery, S3 and its own products. There is no connector for Australian platform files or unit registry data. You get a generic ingestion API instead, which is a pipe, not a connector. Someone has to build each feed, parse each platform’s format, match up adviser and dealer-group codes, handle restated data, and keep it working when formats change.
Salesforce does give you low-code tooling that reduces the effort, and its certifications cut down the audit work — though they do not remove it, because you stay accountable for assets you do not operate. And the feeds still have to be built. With the infrastructure, the agent framework and the governance scaffolding supplied, what is left is the six feeds, the configuration and the handover — about one engineer-year, or roughly $200,000 at Australian rates.
Then someone has to run it. One person, fully loaded, is about $200,000 a year — every year, not just the first.
Option 2: build it yourself
Some firms decide to own it outright. The maths is short. Building the architecture, governance and tooling from scratch is about two engineer-years, and running the stack takes another engineer and a half: $700,000 of people in year one. Add the three platform layers and compliance audits, and you are at $821,000.
The platform lines are the stack you assemble yourself: a warehouse with ingestion, transformation and orchestration around it, Power BI on top, and the gateway, guardrails, vector store and evaluation tooling that serve the models. Salesforce bundles that last layer into its agent product; on your own it is one more thing to run. The $60,000 data line is deliberately conservative — the warehouse alone benchmarks at $28,000 to $50,000, and ingestion and orchestration sit on top of that.
The tooling itself is surprisingly cheap. A production vector store runs a few hundred dollars a month, and evaluation and observability platforms are seat-priced in the tens of dollars. That is the point: the software is not what costs you. The build is a one-off. The people are not. The person who runs it is there every year, and so is the risk that both engineers resign in the same quarter and take the knowledge with them. We have written separately on why an embedded engineer beats assembling that team in-house.
Option 3: buy from us
A fixed platform fee plus what you use. For a 20-seat deployment that is around $160,000 a year.
That price already contains the infrastructure, the data feeds and their upkeep, the compliance audits, and the CRM administration. No implementation fee. No hire.
But the price is not the reason to choose it. These are.
It was built with asset managers, not for them. Every line in that build row — the architecture, the governance framework, the platform and registry feeds, the adviser and dealer-group mapping — we have already done, in production, with real firms. You are buying the result of that work instead of commissioning it.
It is specific to this industry. A generic platform can be pointed at fund flow data. Ours was designed around it, with 35+ prebuilt industry connectors — platforms, registries, custodians, market data. That is why the connector problem in Option 1 mostly does not arise here: the feeds you need are far more likely to already exist than to need building.
It was tested, and a good deal of it was thrown away. Workflows that looked right in a demo and failed in the field got removed. Someone pays for that learning once. We have already paid it.
And we are accountable for whether it gets used — which is a big enough claim that it gets its own section below.
It is also all Australian: every model runs through our own gateway in Sydney and Melbourne, so your data never leaves the country.
How our price is built
The figure in Option 3 is the output of a model, not a list price. Every client is priced off the same rate card. The rates do not change from one firm to the next; the counts do. If two of our clients compared invoices, the rates would match and the counts would not — and both would be able to see why.
The platform fee
What we run for you whether you log in daily or not at all. Charged per module, so you are not billed for capability you never turned on.
- Infrastructure — hosting, the Australian model gateway, backups, monitoring, recoveryper year
- Assurance & audits — controls, access and data handling, the evidence behind them, the annual review cycleper year
- Data foundation — the warehouse, the models, the quality layer underneath everythingper module
- Distribution, operations, investment modules — taken individuallyper module
What you actually use
The lines that move with your firm. Each one is a count you can verify against your own systems, metered on a published unit.
- Integrations — split into small and large, because the effort is not a smooth curveper feed
- Data volume — storage and the processing over itper GB
- AI usage — model consumption, pooled across the firmper M tokens
- Users — the people with an accountper user / month
- Retained capacity — a named share of our team, in whole or half unitsper month
The fixed part covers what runs whether you use it or not, which is why the compliance row above reads “included” — we run that programme once, across every client, though the sign-off stays yours. Modules sit on separate lines so a firm that never opens the operations module is not subsidising it, and there is no implementation fee because there is no implementation project. The variable part is metered on counts you can check against your own systems.
One thing worth knowing before you ever see an invoice: four lines account for nearly all of it — the platform fee, the seats, the integrations and any retained capacity. Every one of them is a person or a piece of work. The model usage itself, the AI everyone is actually asking about, is the smallest line on the page.
Where the money really goes
Look at the US vendor column again. The AI platform is $8,000 out of $486,000. The two people lines are $400,000 — more than four-fifths of the total.
Comparing per-seat prices and token rates compares the rounding error. The question that decides your budget is whether the platform arrives with the people who operate it.
The question cost cannot answer
Everything above is a budget exercise. The question that actually decides whether the money was well spent is simpler, and no table can price it: does anyone use the thing?
Our answer is 100% adoption at Global X and Talaria, who run the platform on an ongoing subscription — not licences issued, people using it. It is a number most AI programmes cannot produce at any size, and it is the only one on this page that says the spending worked.
We measure adoption and we measure impact: what got used, which conversations happened, which flows moved. If it is not being used, that is our problem to solve — not a line item you keep paying for while wondering. Neither of the other two options makes that its problem. Buy the vendor stack and adoption is yours to drive; build it yourself and measurement is one more thing to build.
When we’re not the right answer
Every vendor comparison is built to make the vendor win, so here is where ours stops working.
Everything rests on the people line. Assume you already employ someone with genuinely spare time to run it, and that you need only one or two simple feeds rather than six. At half an administrator and two feeds, the US vendor column falls to roughly $286,000 in year one and $186,000 after it — within about $25,000 of us once the build is behind you. We would not be cheaper.
So: if you have a Salesforce administrator with spare capacity, simple data needs, and nothing you need from outside the CRM, buy the US vendor stack. We would rather say that now than find out together in month six.
If you would have to hire, or the data you need sits in platform files and registry extracts, the comparison is not close.
And there is a fair question back at us. Part of the case against building is that two engineers can resign in the same quarter and take the knowledge with them. Point that at a supplier of our size and it is a reasonable thing to ask. The honest answer is about where the knowledge sits: in an in-house build it lives in a couple of people’s heads; with us it lives in a platform that has to serve every client, which is what forces it to be documented, repeatable and not dependent on any one person — ours or yours. We have made that argument in full elsewhere. It is a better answer than “trust us”, but it is not a complete one. Concentration risk with a small supplier is real, you should price it, and you should make it a term of the contract rather than take our word for it.
Sources
- Agentforce pricing — per conversation, Flex Credits and per-user tiers. Salesforce
- Claude Team and Enterprise seat pricing. Anthropic
- ChatGPT Business and Enterprise pricing, including the Enterprise seat minimum. OpenAI
- Data Cloud entry package list price, $60,000/year. Salesforce Data Cloud
- Available Data Cloud connectors. Salesforce connectors directory
- BI seat pricing — Tableau Cloud at $15 Viewer / $42 Explorer / $75 Creator per user per month. Tableau
- Power BI at $14 Pro / $24 Premium Per User per month. Microsoft
- Salesforce Foundations free tier — 200,000 Flex Credits, Agent Builder and Prompt Builder at no cost on Enterprise Edition and above.
- Vector database pricing — Pinecone from $50/month, $500/month at the enterprise tier; Weaviate comparable. 2026 vector database cost comparisons.
- LLM observability and evaluation tooling, from $29–$39 per seat per month — 2026 LangSmith, Langfuse and Arize pricing comparisons.
- Mid-size warehouse spend, roughly $28,000–$50,000 a year — 2026 Snowflake and Databricks cost benchmarks. A full stack adding ingestion, transformation, BI and orchestration runs higher.
- Third-party certifications reduce but do not remove an entity’s own assurance obligations — APRA CPS 234 guidance.
- Enterprise Salesforce implementation range, $150,000–$500,000+ — 2026 Salesforce partner implementation cost guides.
- Australian data engineer salaries, $150,000 base — Morgan McKinley 2026 salary guide.
- Contractor day rates, $850–$1,150 — Talent and Re:Sourced 2026 rate guides.
- Superannuation guarantee 12% — Australian Taxation Office.
- Payroll tax 5.45% NSW / 4.85% VIC — Revenue NSW and the Victorian State Revenue Office.
- Data residency within Australia — Amazon Bedrock security and privacy.
Figures are illustrative, modelled on a representative 20-seat deployment in its first year, and rounded. Licence rates are 2026 published list prices; salary and contractor rates are 2026 Australian market guides. Your costs will vary with scope, negotiated pricing and how many data feeds you need.