Oded Har-Tal

The limiting factor becomes the data environment AI operates on.

• Har-Tal says Datarails chose early to invest in the platform and financial data environment for AI-powered finance rather than only refining traditional FP&A interfaces.
• The reasoning is that AI can only be as powerful as the financial data layer beneath it, so governed, real-time, traceable data has to come before trustworthy automation.

In one post, Har-Tal contrasts two decades of human-centered FP&A tools with AI that can increasingly perform those tasks itself. Har-Tal then names data infrastructure as the new bottleneck and describes Datarails opening its data layer so finance teams can connect static reporting to live, intelligent execution.

The argument moves from a long company history of helping finance teams work with Excel and insights toward a platform designed for AI to act on the data directly.

Read original · Original 2 · Original 3

The replacement Har-Tal names is the Finance Operating System, which positions Datarails as a platform for the AI of the CFO rather than another tool for CFOs. The stated concrete move is from building products for CFOs to opening a governed, real-time, traceable foundation that AI can use.

Har-Tal writes that FP&A software is dead and announces the Finance Operating System era, saying Datarails is making a bold company bet by opening its data layer to provide a governed, real-time, and fully traceable foundation for AI execution in finance. The post is promotional for Datarails and states a company position, so it is evidence of Har-Tal's public argument rather than independent market proof.

Read original · Original 2

Har-Tal benchmarked MCP against direct API calls and CLI on the same task with the same tools using Claude and GPT-4o.

• The overhead people worry about was negligible in that test: same turns, same latency, same results.
• Har-Tal's nuanced conclusion is that MCP still wins for tool discovery, while direct code composition is becoming stronger for deep, multi-step workflows where agents compose dependent operations.

In the benchmark post, Har-Tal describes an agent that needs to look up an account, use that result to fetch transactions, and then compute a balance.

• With MCP, each step round-trips through the model; with direct function calls, the chain can happen in one call.
• That concrete workflow is the basis for Har-Tal's recommendation by use case.

The post reframes the debate away from performance overhead and toward workflow shape: protocol where discovery matters, direct code where deep multi-step composition matters.

Read original

Skills inject knowledge into the current context.

• Sub-agents delegate to a child agent that works in isolation and returns a summary.
• Har-Tal argues the real cost is not tokens per task but context pollution: every file read, tool result, and intermediate step stays in the conversation with skills.
• After a few tasks, that can leave far more tokens in context, making later API calls more expensive, degrading model performance through attention dilution, and hitting context limits faster.
• The rule of thumb is simple task to skill, complex or multi-step task to sub-agent.

Har-Tal says that after just three tasks, the skill pattern can leave 50x more tokens in context than sub-agents. To show the difference, Har-Tal built a self-contained interactive notebook with live Claude API calls, real token measurements, and flow diagrams for each pattern.

Read original

The Datarails retrospective starts from the belief that finance teams should not have to choose between Excel and real insights, then describes an end-to-end platform for the CFO office with AI at its core.

• Har-Tal credits Didi Gurfinkel and Eyal Cohen for building Datarails together and welcomes One Peak as a partner.
• Those names are attributions from Har-Tal's post, not separate guest arguments.

In the 11-year retrospective, Har-Tal says watching AI move from an idea to something that helps CFOs close faster, understand cash, and trust their numbers has been one of the most rewarding parts of the journey. That milestone framing reinforces the same platform thesis: AI becomes useful when it sits on an operating foundation the CFO can rely on.

Read original · Original 2 · Original 3

It answers what Oded Har-Tal thinks is changing in FP&A and why data infrastructure, not another dashboard, is the center of that change.

Oded Har-Tal argues that finance software is moving from static reporting toward a governed, real-time data platform that AI can act on.

• The post says Datarails is opening its data layer so finance teams can connect legacy reporting to live execution, and it frames the shift as a new category: the Finance Operating System.
• The concrete move is from building products for CFOs to building the platform for the AI of the CFO.

For brands planning a campaign

A co-authored CFO guest post can use this thesis as its spine: contrast static reporting with live, traceable execution, then require governance and data readiness rather than generic AI enthusiasm. The category language also gives the article a natural headline and structure.

A newcomer gets the argument behind the FinanceOS positioning in a short, direct form.

Oded Har-Tal says AI will change finance software more than most people expected, but the bottleneck is shifting from human task work to the financial data layer.

• The post argues that AI can only be as powerful as the data environment it operates on, and that Datarails chose to invest in that environment rather than only refining traditional FP&A interfaces.
• It ends by asking whether finance teams are actually ready for the transition.

For brands planning a campaign

For a CFO audience, the co-authored piece can be built as a readiness argument: what financial data must be governed, traceable, and real-time before AI can help with closing, cash, and forecasting. The closing question also provides a ready-made reader engagement hook.

It shows Oded Har-Tal speaking as someone who has worked through the Excel-to-platform transition, not only as a commentator on AI trends.

Oded Har-Tal looks back on 11 years of building Datarails and describes the company's move from a belief that finance teams should not have to choose between Excel and real insights to an end-to-end CFO office platform with AI at its core. The post says AI is helping CFOs close faster, understand cash, and trust their numbers, thanks Didi Gurfinkel and Eyal Cohen for building it together, and welcomes One Peak as a partner.

For brands planning a campaign

Use this as credibility context for a guest post rather than making funding the story. The Excel-versus-insights tension is a useful opening scene, and the operating history supports a co-authored article about what CFO offices need as AI moves from idea to daily use.

Start here to see how they evaluate AI infrastructure claims with benchmarks rather than assertions.

Oded Har-Tal benchmarked MCP, direct API, and CLI on the same task with the same tools using Claude and GPT-4o.

• The post reports that the overhead people worry about was negligible: same turns, same latency, same results.
• The real issue, in this account, is that agents increasingly compose dependent operations, and with MCP each step round-trips through the model.
• MCP still wins for tool discovery, but direct code composition is becoming stronger for deep, multi-step workflows.

For brands planning a campaign

If the campaign also targets developer or technical readers, this supplies a concrete angle on agent architecture choices. For a CFO-focused brief, it is better used as an optional sidebar showing that Oded Har-Tal tests claims before making recommendations.

It answers a specific builder question and shows Oded Har-Tal using measured examples and a runnable notebook, not just opinion.

Oded Har-Tal compares two ways to extend AI coding assistants: skills inject knowledge into the current context, while sub-agents delegate to an isolated child agent that returns a summary.

• The post says the real cost is not tokens per task but context pollution: after just three tasks, the skill pattern can leave 50x more tokens in context, making later API calls more expensive, degrading model performance through attention dilution, and hitting context limits faster.
• The rule of thumb is simple task to skill, complex or multi-step task to sub-agent.
• The post points to an interactive notebook with live Claude API calls, real token measurements, and flow diagrams.

For brands planning a campaign

This demonstrates the kind of hands-on artifact that can strengthen a co-authored article: a tested framework, a clear rule, and a resource readers can run themselves. If the B2B SaaS campaign wants a practical companion asset, this is evidence Oded Har-Tal can produce that style of explainer.

In one post, Har-Tal says Datarails chose early to invest in the platform and financial data environment for AI-powered finance rather than only refining traditional FP&A interfaces. In another, Har-Tal describes Datarails opening a governed, real-time, traceable foundation so finance teams can move from static reporting to live, intelligent execution.

Har-Tal explains the shift by contrasting two decades of human-centered FP&A tools with AI that can perform core tasks itself, then names data infrastructure as the new bottleneck.

For brands planning a campaign

For a B2B SaaS company selling to CFOs, this is the strongest available guest-post angle: a co-authored article on why AI in finance requires governed, traceable financial data before automation can be trusted. Use Har-Tal's framing as the argument and add the SaaS company's specific workflow evidence.

Examples: Source 1 · Source 2 · Source 3

The post says the company is opening its data layer to provide a governed, real-time, traceable foundation for AI execution.

Category reframing: declare the old category obsolete, name a new one, and connect the new category to a platform capability.

The post is promotional for Datarails and states a company bet. Treat it as evidence of Har-Tal's public positioning, not independent market proof.

For brands planning a campaign

A guest post can use this category-shift argument to challenge legacy planning suites and propose a CFO-ready operating model. The claim is bold; a brand should pair it with concrete implementation detail rather than repeat the slogan.

Examples: Source 1

Har-Tal says the overhead people worry about was negligible and argues that MCP still wins for tool discovery while direct code composition is becoming stronger for deep, multi-step workflows.

Controlled comparison: same task, same tools, multiple models, then a nuanced recommendation by use case.

This post is about AI developer tooling, not finance buyers, so it supports credibility in technical evaluation more than direct CFO relevance.

For brands planning a campaign

This shows a hands-on evaluator who tests claims before recommending an approach. For a CFO-oriented guest post, translate the same discipline into finance workflows: show a tested comparison rather than asserting that AI improves a process.

Examples: Source 1

Har-Tal argues the real cost is context pollution, not tokens per task, and notes the skill pattern can leave far more tokens in context after a few tasks. Har-Tal also built a self-contained notebook with live API calls and token measurements.

Rule of thumb plus interactive demonstration: simple task use a skill; complex or multi-step task use a sub-agent.

For brands planning a campaign

For a finance-audience guest post, this is a secondary technical example. It can support a section on managing AI cost and reliability, but the brand should not lead with developer tooling when the buyer is a CFO.

Examples: Source 1

Milestone reflection that thanks co-founders and names the investor, then connects the company story to the product thesis.

For brands planning a campaign

A co-authored article can draw on Har-Tal's long company-building perspective and investor relationship, but the brand should keep the focus on the reader's finance problem rather than a Datarails milestone announcement.

Examples: Source 1

He speaks for the company on its product direction (FinanceOS) and its AI-and-finance thesis, and credits Didi Gurfinkel and Eyal Cohen as people who built it with him.

View example

The connection described is between the fund and Datarails, not a personal sponsorship, endorsement, or client engagement for Har-Tal.

View example

Both write about AI copilots assisting controllers, FP&A, and CFOs, and both frame the shift as a change in what the finance function does rather than a feature upgrade. Both also write for a finance-leader audience rather than a developer one.

Peddi stays at the level of the category shift, describing a move from record-to-report toward predict-to-act and from ERP toward a finance decision layer, without naming a specific data foundation the way Har-Tal does.

• Har-Tal's version is anchored in a platform argument about a governed, traceable data layer, and Har-Tal also writes about AI infrastructure outside finance.
• Peddi's posts are denser with bulleted vision statements and hashtags; Har-Tal's are more often built around a single comparison or benchmark.

Read original · Original 2

Both use the what happened, why did it happen, what should I do about it progression, and both describe AI features that explain performance and recommend actions rather than only reporting numbers. Both also tie AI in finance to governance and control requirements.

Chess grounds the argument in observed behavior, reporting what NetSuite customers actually asked across thousands of conversations after a product launched, and announcing a specific skill for Claude and ChatGPT.

• Har-Tal more often argues category and platform positioning, and his comparison material sits in developer tooling rather than in finance product usage data.
• If you want customer-behavior evidence, read Chess; if you want the platform thesis behind it, read Har-Tal.

Read original · Original 2

Both use before-and-after framing about what finance teams did historically and what they can do now, and both put concrete numbers behind the argument rather than leaving it abstract.

Osburn's posts are vendor-anchored and quantitative, built around time saved, accuracy rates, and ticket reductions for a finance-operations audience, and they read as product-positioning pieces.

• Har-Tal writes more about the layer AI sits on and about how to choose between architectures, and tests claims in a benchmark rather than citing an operational benchmark.
• Osburn argues the change is happening; Har-Tal argues about the infrastructure required for it to be trustworthy.

Read original · Original 2

The two reviewed posts argue that ERP tells you what happened but does not help you decide what to do next, that decision latency is now a form of financial risk, and that CFOs are shifting from record-to-report toward predict-to-act, continuous finance, and a "finance decision layer" where revenue leakage is caught before recognition, cash flow is optimized dynamically, and forecasts self-adjust on operational signals.

• For a CFO audience that framing gives the guest post a sharper opening question — is your ERP still reporting history, or predicting tomorrow? — and keeps the piece on operating cadence instead of platform architecture.
• Peddi also writes in short, list-driven blocks that translate cleanly into pull quotes and section headers.

The tradeoff is evidential depth.

• Neither reviewed post reports a test, benchmark, measurement, or named deployment; the argument runs on assertions about market direction and a vision of autonomous decisioning.
• Oded Har-Tal's reviewed work, by contrast, includes a controlled MCP-versus-direct-API-versus-CLI benchmark and token measurements for skills versus sub-agents, so a partner choosing Peddi trades that tested-comparison habit for a stronger forward-looking narrative.
• A CFO article leaning on Peddi would need the commissioning brand to supply its own customer proof, because the posts do not carry it.
• The reviewed material names no employer or product, and no audience figures or commercial terms were supplied, so reach and availability cannot be assessed.

Examples: Source 1 · Source 2

One reviewed post reports what NetSuite customers actually ask Ask Oracle — which invoices are overdue, what is driving a cash shortage, which costs are pressuring margins, which anomalies deserve attention before close — and traces the progression from what happened to why it happened to what deserves attention to what to do about it.

• The other announces the NetSuite Finance Analyst Skill for Claude and ChatGPT and lists its guardrails: authoritative NetSuite financial reports as the source, standard finance workflows such as close management and variance analysis, and controls around period status, materiality, and source authority.
• For a CFO reader, that supplies concrete examples of AI inside a finance stack plus a working vocabulary of controls, which is harder to find than general AI-in-finance commentary.

The tradeoff is that the evidence is single-vendor and product-shaped.

• Both posts sit inside one company's AI rollout, so an article built on them risks reading as a product announcement, and the claims describe that vendor's customers rather than a cross-platform readiness test.
• Chess does not offer the platform-swap argument — that the bottleneck has moved to the financial data layer — which is exactly what makes Har-Tal's reviewed framing useful to a SaaS company selling into the CFO office.
• The reviewed material contains no audience numbers, no format or fee terms, and no confirmation of the named role beyond the posts themselves.

Examples: Source 1 · Source 2

The reviewed posts carry concrete figures: reconciliation consuming up to 600 hours a year for the average company, 330 hours a year saved on expense processing at a 200-person company, documents captured at 93% accuracy, expense submissions 60% faster, and 40% fewer finance support tickets.

• Both frame the change as finance moving from transaction processor to real-time oversight, with AI agents collecting receipts, validating transactions, applying policy, and reconciling continuously in the background so teams close faster with more visibility into spend.
• That is useful for the part of the article that has to show a CFO what changes in hours and headcount terms, and the numbers convert easily into a short, scannable comparison.

The tradeoff is that the figures are vendor claims tied to one product, Payhawk's AI Office of the CFO, with no methodology, sample, or independent source in the reviewed text, so a co-authored article would have to attribute them carefully or replace them with the commissioning brand's own data.

• The focus is also narrower than Har-Tal's: spend, expenses, and close cleanup rather than the wider data-infrastructure question about what AI needs underneath it.
• Osburn complements the chosen creator's spine instead of replacing it.
• No audience figures, commercial terms, or confirmation of the named role beyond the posts were supplied.

Examples: Source 1 · Source 2